David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 1 | ; This file is generated from a similarly-named Perl script in the BoringSSL |
| 2 | ; source tree. Do not edit by hand. |
| 3 | |
| 4 | %ifidn __OUTPUT_FORMAT__, win64 |
| 5 | default rel |
| 6 | %define XMMWORD |
| 7 | %define YMMWORD |
| 8 | %define ZMMWORD |
| 9 | %define _CET_ENDBR |
| 10 | |
| 11 | %ifdef BORINGSSL_PREFIX |
| 12 | %include "boringssl_prefix_symbols_nasm.inc" |
| 13 | %endif |
| 14 | section .text code align=64 |
| 15 | |
| 16 | |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 17 | global bn_mul_mont_gather5_nohw |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 18 | |
| 19 | ALIGN 64 |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 20 | bn_mul_mont_gather5_nohw: |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 21 | mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 22 | mov QWORD[16+rsp],rsi |
| 23 | mov rax,rsp |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 24 | $L$SEH_begin_bn_mul_mont_gather5_nohw: |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 25 | mov rdi,rcx |
| 26 | mov rsi,rdx |
| 27 | mov rdx,r8 |
| 28 | mov rcx,r9 |
| 29 | mov r8,QWORD[40+rsp] |
| 30 | mov r9,QWORD[48+rsp] |
| 31 | |
| 32 | |
| 33 | |
| 34 | _CET_ENDBR |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 35 | |
| 36 | |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 37 | mov r9d,r9d |
| 38 | mov rax,rsp |
| 39 | |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 40 | movd xmm5,DWORD[56+rsp] |
| 41 | push rbx |
| 42 | |
| 43 | push rbp |
| 44 | |
| 45 | push r12 |
| 46 | |
| 47 | push r13 |
| 48 | |
| 49 | push r14 |
| 50 | |
| 51 | push r15 |
| 52 | |
| 53 | |
| 54 | neg r9 |
| 55 | mov r11,rsp |
| 56 | lea r10,[((-280))+r9*8+rsp] |
| 57 | neg r9 |
| 58 | and r10,-1024 |
| 59 | |
| 60 | |
| 61 | |
| 62 | |
| 63 | |
| 64 | |
| 65 | |
| 66 | |
| 67 | |
| 68 | sub r11,r10 |
| 69 | and r11,-4096 |
| 70 | lea rsp,[r11*1+r10] |
| 71 | mov r11,QWORD[rsp] |
| 72 | cmp rsp,r10 |
| 73 | ja NEAR $L$mul_page_walk |
| 74 | jmp NEAR $L$mul_page_walk_done |
| 75 | |
| 76 | $L$mul_page_walk: |
| 77 | lea rsp,[((-4096))+rsp] |
| 78 | mov r11,QWORD[rsp] |
| 79 | cmp rsp,r10 |
| 80 | ja NEAR $L$mul_page_walk |
| 81 | $L$mul_page_walk_done: |
| 82 | |
| 83 | lea r10,[$L$inc] |
| 84 | mov QWORD[8+r9*8+rsp],rax |
| 85 | |
| 86 | $L$mul_body: |
| 87 | |
| 88 | lea r12,[128+rdx] |
| 89 | movdqa xmm0,XMMWORD[r10] |
| 90 | movdqa xmm1,XMMWORD[16+r10] |
| 91 | lea r10,[((24-112))+r9*8+rsp] |
| 92 | and r10,-16 |
| 93 | |
| 94 | pshufd xmm5,xmm5,0 |
| 95 | movdqa xmm4,xmm1 |
| 96 | movdqa xmm2,xmm1 |
| 97 | paddd xmm1,xmm0 |
| 98 | pcmpeqd xmm0,xmm5 |
| 99 | DB 0x67 |
| 100 | movdqa xmm3,xmm4 |
| 101 | paddd xmm2,xmm1 |
| 102 | pcmpeqd xmm1,xmm5 |
| 103 | movdqa XMMWORD[112+r10],xmm0 |
| 104 | movdqa xmm0,xmm4 |
| 105 | |
| 106 | paddd xmm3,xmm2 |
| 107 | pcmpeqd xmm2,xmm5 |
| 108 | movdqa XMMWORD[128+r10],xmm1 |
| 109 | movdqa xmm1,xmm4 |
| 110 | |
| 111 | paddd xmm0,xmm3 |
| 112 | pcmpeqd xmm3,xmm5 |
| 113 | movdqa XMMWORD[144+r10],xmm2 |
| 114 | movdqa xmm2,xmm4 |
| 115 | |
| 116 | paddd xmm1,xmm0 |
| 117 | pcmpeqd xmm0,xmm5 |
| 118 | movdqa XMMWORD[160+r10],xmm3 |
| 119 | movdqa xmm3,xmm4 |
| 120 | paddd xmm2,xmm1 |
| 121 | pcmpeqd xmm1,xmm5 |
| 122 | movdqa XMMWORD[176+r10],xmm0 |
| 123 | movdqa xmm0,xmm4 |
| 124 | |
| 125 | paddd xmm3,xmm2 |
| 126 | pcmpeqd xmm2,xmm5 |
| 127 | movdqa XMMWORD[192+r10],xmm1 |
| 128 | movdqa xmm1,xmm4 |
| 129 | |
| 130 | paddd xmm0,xmm3 |
| 131 | pcmpeqd xmm3,xmm5 |
| 132 | movdqa XMMWORD[208+r10],xmm2 |
| 133 | movdqa xmm2,xmm4 |
| 134 | |
| 135 | paddd xmm1,xmm0 |
| 136 | pcmpeqd xmm0,xmm5 |
| 137 | movdqa XMMWORD[224+r10],xmm3 |
| 138 | movdqa xmm3,xmm4 |
| 139 | paddd xmm2,xmm1 |
| 140 | pcmpeqd xmm1,xmm5 |
| 141 | movdqa XMMWORD[240+r10],xmm0 |
| 142 | movdqa xmm0,xmm4 |
| 143 | |
| 144 | paddd xmm3,xmm2 |
| 145 | pcmpeqd xmm2,xmm5 |
| 146 | movdqa XMMWORD[256+r10],xmm1 |
| 147 | movdqa xmm1,xmm4 |
| 148 | |
| 149 | paddd xmm0,xmm3 |
| 150 | pcmpeqd xmm3,xmm5 |
| 151 | movdqa XMMWORD[272+r10],xmm2 |
| 152 | movdqa xmm2,xmm4 |
| 153 | |
| 154 | paddd xmm1,xmm0 |
| 155 | pcmpeqd xmm0,xmm5 |
| 156 | movdqa XMMWORD[288+r10],xmm3 |
| 157 | movdqa xmm3,xmm4 |
| 158 | paddd xmm2,xmm1 |
| 159 | pcmpeqd xmm1,xmm5 |
| 160 | movdqa XMMWORD[304+r10],xmm0 |
| 161 | |
| 162 | paddd xmm3,xmm2 |
| 163 | DB 0x67 |
| 164 | pcmpeqd xmm2,xmm5 |
| 165 | movdqa XMMWORD[320+r10],xmm1 |
| 166 | |
| 167 | pcmpeqd xmm3,xmm5 |
| 168 | movdqa XMMWORD[336+r10],xmm2 |
| 169 | pand xmm0,XMMWORD[64+r12] |
| 170 | |
| 171 | pand xmm1,XMMWORD[80+r12] |
| 172 | pand xmm2,XMMWORD[96+r12] |
| 173 | movdqa XMMWORD[352+r10],xmm3 |
| 174 | pand xmm3,XMMWORD[112+r12] |
| 175 | por xmm0,xmm2 |
| 176 | por xmm1,xmm3 |
| 177 | movdqa xmm4,XMMWORD[((-128))+r12] |
| 178 | movdqa xmm5,XMMWORD[((-112))+r12] |
| 179 | movdqa xmm2,XMMWORD[((-96))+r12] |
| 180 | pand xmm4,XMMWORD[112+r10] |
| 181 | movdqa xmm3,XMMWORD[((-80))+r12] |
| 182 | pand xmm5,XMMWORD[128+r10] |
| 183 | por xmm0,xmm4 |
| 184 | pand xmm2,XMMWORD[144+r10] |
| 185 | por xmm1,xmm5 |
| 186 | pand xmm3,XMMWORD[160+r10] |
| 187 | por xmm0,xmm2 |
| 188 | por xmm1,xmm3 |
| 189 | movdqa xmm4,XMMWORD[((-64))+r12] |
| 190 | movdqa xmm5,XMMWORD[((-48))+r12] |
| 191 | movdqa xmm2,XMMWORD[((-32))+r12] |
| 192 | pand xmm4,XMMWORD[176+r10] |
| 193 | movdqa xmm3,XMMWORD[((-16))+r12] |
| 194 | pand xmm5,XMMWORD[192+r10] |
| 195 | por xmm0,xmm4 |
| 196 | pand xmm2,XMMWORD[208+r10] |
| 197 | por xmm1,xmm5 |
| 198 | pand xmm3,XMMWORD[224+r10] |
| 199 | por xmm0,xmm2 |
| 200 | por xmm1,xmm3 |
| 201 | movdqa xmm4,XMMWORD[r12] |
| 202 | movdqa xmm5,XMMWORD[16+r12] |
| 203 | movdqa xmm2,XMMWORD[32+r12] |
| 204 | pand xmm4,XMMWORD[240+r10] |
| 205 | movdqa xmm3,XMMWORD[48+r12] |
| 206 | pand xmm5,XMMWORD[256+r10] |
| 207 | por xmm0,xmm4 |
| 208 | pand xmm2,XMMWORD[272+r10] |
| 209 | por xmm1,xmm5 |
| 210 | pand xmm3,XMMWORD[288+r10] |
| 211 | por xmm0,xmm2 |
| 212 | por xmm1,xmm3 |
| 213 | por xmm0,xmm1 |
| 214 | |
| 215 | pshufd xmm1,xmm0,0x4e |
| 216 | por xmm0,xmm1 |
| 217 | lea r12,[256+r12] |
| 218 | DB 102,72,15,126,195 |
| 219 | |
| 220 | mov r8,QWORD[r8] |
| 221 | mov rax,QWORD[rsi] |
| 222 | |
| 223 | xor r14,r14 |
| 224 | xor r15,r15 |
| 225 | |
| 226 | mov rbp,r8 |
| 227 | mul rbx |
| 228 | mov r10,rax |
| 229 | mov rax,QWORD[rcx] |
| 230 | |
| 231 | imul rbp,r10 |
| 232 | mov r11,rdx |
| 233 | |
| 234 | mul rbp |
| 235 | add r10,rax |
| 236 | mov rax,QWORD[8+rsi] |
| 237 | adc rdx,0 |
| 238 | mov r13,rdx |
| 239 | |
| 240 | lea r15,[1+r15] |
| 241 | jmp NEAR $L$1st_enter |
| 242 | |
| 243 | ALIGN 16 |
| 244 | $L$1st: |
| 245 | add r13,rax |
| 246 | mov rax,QWORD[r15*8+rsi] |
| 247 | adc rdx,0 |
| 248 | add r13,r11 |
| 249 | mov r11,r10 |
| 250 | adc rdx,0 |
| 251 | mov QWORD[((-16))+r15*8+rsp],r13 |
| 252 | mov r13,rdx |
| 253 | |
| 254 | $L$1st_enter: |
| 255 | mul rbx |
| 256 | add r11,rax |
| 257 | mov rax,QWORD[r15*8+rcx] |
| 258 | adc rdx,0 |
| 259 | lea r15,[1+r15] |
| 260 | mov r10,rdx |
| 261 | |
| 262 | mul rbp |
| 263 | cmp r15,r9 |
| 264 | jne NEAR $L$1st |
| 265 | |
| 266 | |
| 267 | add r13,rax |
| 268 | adc rdx,0 |
| 269 | add r13,r11 |
| 270 | adc rdx,0 |
| 271 | mov QWORD[((-16))+r9*8+rsp],r13 |
| 272 | mov r13,rdx |
| 273 | mov r11,r10 |
| 274 | |
| 275 | xor rdx,rdx |
| 276 | add r13,r11 |
| 277 | adc rdx,0 |
| 278 | mov QWORD[((-8))+r9*8+rsp],r13 |
| 279 | mov QWORD[r9*8+rsp],rdx |
| 280 | |
| 281 | lea r14,[1+r14] |
| 282 | jmp NEAR $L$outer |
| 283 | ALIGN 16 |
| 284 | $L$outer: |
| 285 | lea rdx,[((24+128))+r9*8+rsp] |
| 286 | and rdx,-16 |
| 287 | pxor xmm4,xmm4 |
| 288 | pxor xmm5,xmm5 |
| 289 | movdqa xmm0,XMMWORD[((-128))+r12] |
| 290 | movdqa xmm1,XMMWORD[((-112))+r12] |
| 291 | movdqa xmm2,XMMWORD[((-96))+r12] |
| 292 | movdqa xmm3,XMMWORD[((-80))+r12] |
| 293 | pand xmm0,XMMWORD[((-128))+rdx] |
| 294 | pand xmm1,XMMWORD[((-112))+rdx] |
| 295 | por xmm4,xmm0 |
| 296 | pand xmm2,XMMWORD[((-96))+rdx] |
| 297 | por xmm5,xmm1 |
| 298 | pand xmm3,XMMWORD[((-80))+rdx] |
| 299 | por xmm4,xmm2 |
| 300 | por xmm5,xmm3 |
| 301 | movdqa xmm0,XMMWORD[((-64))+r12] |
| 302 | movdqa xmm1,XMMWORD[((-48))+r12] |
| 303 | movdqa xmm2,XMMWORD[((-32))+r12] |
| 304 | movdqa xmm3,XMMWORD[((-16))+r12] |
| 305 | pand xmm0,XMMWORD[((-64))+rdx] |
| 306 | pand xmm1,XMMWORD[((-48))+rdx] |
| 307 | por xmm4,xmm0 |
| 308 | pand xmm2,XMMWORD[((-32))+rdx] |
| 309 | por xmm5,xmm1 |
| 310 | pand xmm3,XMMWORD[((-16))+rdx] |
| 311 | por xmm4,xmm2 |
| 312 | por xmm5,xmm3 |
| 313 | movdqa xmm0,XMMWORD[r12] |
| 314 | movdqa xmm1,XMMWORD[16+r12] |
| 315 | movdqa xmm2,XMMWORD[32+r12] |
| 316 | movdqa xmm3,XMMWORD[48+r12] |
| 317 | pand xmm0,XMMWORD[rdx] |
| 318 | pand xmm1,XMMWORD[16+rdx] |
| 319 | por xmm4,xmm0 |
| 320 | pand xmm2,XMMWORD[32+rdx] |
| 321 | por xmm5,xmm1 |
| 322 | pand xmm3,XMMWORD[48+rdx] |
| 323 | por xmm4,xmm2 |
| 324 | por xmm5,xmm3 |
| 325 | movdqa xmm0,XMMWORD[64+r12] |
| 326 | movdqa xmm1,XMMWORD[80+r12] |
| 327 | movdqa xmm2,XMMWORD[96+r12] |
| 328 | movdqa xmm3,XMMWORD[112+r12] |
| 329 | pand xmm0,XMMWORD[64+rdx] |
| 330 | pand xmm1,XMMWORD[80+rdx] |
| 331 | por xmm4,xmm0 |
| 332 | pand xmm2,XMMWORD[96+rdx] |
| 333 | por xmm5,xmm1 |
| 334 | pand xmm3,XMMWORD[112+rdx] |
| 335 | por xmm4,xmm2 |
| 336 | por xmm5,xmm3 |
| 337 | por xmm4,xmm5 |
| 338 | |
| 339 | pshufd xmm0,xmm4,0x4e |
| 340 | por xmm0,xmm4 |
| 341 | lea r12,[256+r12] |
| 342 | |
| 343 | mov rax,QWORD[rsi] |
| 344 | DB 102,72,15,126,195 |
| 345 | |
| 346 | xor r15,r15 |
| 347 | mov rbp,r8 |
| 348 | mov r10,QWORD[rsp] |
| 349 | |
| 350 | mul rbx |
| 351 | add r10,rax |
| 352 | mov rax,QWORD[rcx] |
| 353 | adc rdx,0 |
| 354 | |
| 355 | imul rbp,r10 |
| 356 | mov r11,rdx |
| 357 | |
| 358 | mul rbp |
| 359 | add r10,rax |
| 360 | mov rax,QWORD[8+rsi] |
| 361 | adc rdx,0 |
| 362 | mov r10,QWORD[8+rsp] |
| 363 | mov r13,rdx |
| 364 | |
| 365 | lea r15,[1+r15] |
| 366 | jmp NEAR $L$inner_enter |
| 367 | |
| 368 | ALIGN 16 |
| 369 | $L$inner: |
| 370 | add r13,rax |
| 371 | mov rax,QWORD[r15*8+rsi] |
| 372 | adc rdx,0 |
| 373 | add r13,r10 |
| 374 | mov r10,QWORD[r15*8+rsp] |
| 375 | adc rdx,0 |
| 376 | mov QWORD[((-16))+r15*8+rsp],r13 |
| 377 | mov r13,rdx |
| 378 | |
| 379 | $L$inner_enter: |
| 380 | mul rbx |
| 381 | add r11,rax |
| 382 | mov rax,QWORD[r15*8+rcx] |
| 383 | adc rdx,0 |
| 384 | add r10,r11 |
| 385 | mov r11,rdx |
| 386 | adc r11,0 |
| 387 | lea r15,[1+r15] |
| 388 | |
| 389 | mul rbp |
| 390 | cmp r15,r9 |
| 391 | jne NEAR $L$inner |
| 392 | |
| 393 | add r13,rax |
| 394 | adc rdx,0 |
| 395 | add r13,r10 |
| 396 | mov r10,QWORD[r9*8+rsp] |
| 397 | adc rdx,0 |
| 398 | mov QWORD[((-16))+r9*8+rsp],r13 |
| 399 | mov r13,rdx |
| 400 | |
| 401 | xor rdx,rdx |
| 402 | add r13,r11 |
| 403 | adc rdx,0 |
| 404 | add r13,r10 |
| 405 | adc rdx,0 |
| 406 | mov QWORD[((-8))+r9*8+rsp],r13 |
| 407 | mov QWORD[r9*8+rsp],rdx |
| 408 | |
| 409 | lea r14,[1+r14] |
| 410 | cmp r14,r9 |
| 411 | jb NEAR $L$outer |
| 412 | |
| 413 | xor r14,r14 |
| 414 | mov rax,QWORD[rsp] |
| 415 | lea rsi,[rsp] |
| 416 | mov r15,r9 |
| 417 | jmp NEAR $L$sub |
| 418 | ALIGN 16 |
| 419 | $L$sub: sbb rax,QWORD[r14*8+rcx] |
| 420 | mov QWORD[r14*8+rdi],rax |
| 421 | mov rax,QWORD[8+r14*8+rsi] |
| 422 | lea r14,[1+r14] |
| 423 | dec r15 |
| 424 | jnz NEAR $L$sub |
| 425 | |
| 426 | sbb rax,0 |
| 427 | mov rbx,-1 |
| 428 | xor rbx,rax |
| 429 | xor r14,r14 |
| 430 | mov r15,r9 |
| 431 | |
| 432 | $L$copy: |
| 433 | mov rcx,QWORD[r14*8+rdi] |
| 434 | mov rdx,QWORD[r14*8+rsp] |
| 435 | and rcx,rbx |
| 436 | and rdx,rax |
| 437 | mov QWORD[r14*8+rsp],r14 |
| 438 | or rdx,rcx |
| 439 | mov QWORD[r14*8+rdi],rdx |
| 440 | lea r14,[1+r14] |
| 441 | sub r15,1 |
| 442 | jnz NEAR $L$copy |
| 443 | |
| 444 | mov rsi,QWORD[8+r9*8+rsp] |
| 445 | |
| 446 | mov rax,1 |
| 447 | |
| 448 | mov r15,QWORD[((-48))+rsi] |
| 449 | |
| 450 | mov r14,QWORD[((-40))+rsi] |
| 451 | |
| 452 | mov r13,QWORD[((-32))+rsi] |
| 453 | |
| 454 | mov r12,QWORD[((-24))+rsi] |
| 455 | |
| 456 | mov rbp,QWORD[((-16))+rsi] |
| 457 | |
| 458 | mov rbx,QWORD[((-8))+rsi] |
| 459 | |
| 460 | lea rsp,[rsi] |
| 461 | |
| 462 | $L$mul_epilogue: |
| 463 | mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 464 | mov rsi,QWORD[16+rsp] |
| 465 | ret |
| 466 | |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 467 | $L$SEH_end_bn_mul_mont_gather5_nohw: |
| 468 | global bn_mul4x_mont_gather5 |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 469 | |
| 470 | ALIGN 32 |
| 471 | bn_mul4x_mont_gather5: |
| 472 | mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 473 | mov QWORD[16+rsp],rsi |
| 474 | mov rax,rsp |
| 475 | $L$SEH_begin_bn_mul4x_mont_gather5: |
| 476 | mov rdi,rcx |
| 477 | mov rsi,rdx |
| 478 | mov rdx,r8 |
| 479 | mov rcx,r9 |
| 480 | mov r8,QWORD[40+rsp] |
| 481 | mov r9,QWORD[48+rsp] |
| 482 | |
| 483 | |
| 484 | |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 485 | _CET_ENDBR |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 486 | DB 0x67 |
| 487 | mov rax,rsp |
| 488 | |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 489 | push rbx |
| 490 | |
| 491 | push rbp |
| 492 | |
| 493 | push r12 |
| 494 | |
| 495 | push r13 |
| 496 | |
| 497 | push r14 |
| 498 | |
| 499 | push r15 |
| 500 | |
| 501 | $L$mul4x_prologue: |
| 502 | |
| 503 | DB 0x67 |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 504 | |
| 505 | |
| 506 | |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 507 | shl r9d,3 |
| 508 | lea r10,[r9*2+r9] |
| 509 | neg r9 |
| 510 | |
| 511 | |
| 512 | |
| 513 | |
| 514 | |
| 515 | |
| 516 | |
| 517 | |
| 518 | |
| 519 | |
| 520 | lea r11,[((-320))+r9*2+rsp] |
| 521 | mov rbp,rsp |
| 522 | sub r11,rdi |
| 523 | and r11,4095 |
| 524 | cmp r10,r11 |
| 525 | jb NEAR $L$mul4xsp_alt |
| 526 | sub rbp,r11 |
| 527 | lea rbp,[((-320))+r9*2+rbp] |
| 528 | jmp NEAR $L$mul4xsp_done |
| 529 | |
| 530 | ALIGN 32 |
| 531 | $L$mul4xsp_alt: |
| 532 | lea r10,[((4096-320))+r9*2] |
| 533 | lea rbp,[((-320))+r9*2+rbp] |
| 534 | sub r11,r10 |
| 535 | mov r10,0 |
| 536 | cmovc r11,r10 |
| 537 | sub rbp,r11 |
| 538 | $L$mul4xsp_done: |
| 539 | and rbp,-64 |
| 540 | mov r11,rsp |
| 541 | sub r11,rbp |
| 542 | and r11,-4096 |
| 543 | lea rsp,[rbp*1+r11] |
| 544 | mov r10,QWORD[rsp] |
| 545 | cmp rsp,rbp |
| 546 | ja NEAR $L$mul4x_page_walk |
| 547 | jmp NEAR $L$mul4x_page_walk_done |
| 548 | |
| 549 | $L$mul4x_page_walk: |
| 550 | lea rsp,[((-4096))+rsp] |
| 551 | mov r10,QWORD[rsp] |
| 552 | cmp rsp,rbp |
| 553 | ja NEAR $L$mul4x_page_walk |
| 554 | $L$mul4x_page_walk_done: |
| 555 | |
| 556 | neg r9 |
| 557 | |
| 558 | mov QWORD[40+rsp],rax |
| 559 | |
| 560 | $L$mul4x_body: |
| 561 | |
| 562 | call mul4x_internal |
| 563 | |
| 564 | mov rsi,QWORD[40+rsp] |
| 565 | |
| 566 | mov rax,1 |
| 567 | |
| 568 | mov r15,QWORD[((-48))+rsi] |
| 569 | |
| 570 | mov r14,QWORD[((-40))+rsi] |
| 571 | |
| 572 | mov r13,QWORD[((-32))+rsi] |
| 573 | |
| 574 | mov r12,QWORD[((-24))+rsi] |
| 575 | |
| 576 | mov rbp,QWORD[((-16))+rsi] |
| 577 | |
| 578 | mov rbx,QWORD[((-8))+rsi] |
| 579 | |
| 580 | lea rsp,[rsi] |
| 581 | |
| 582 | $L$mul4x_epilogue: |
| 583 | mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 584 | mov rsi,QWORD[16+rsp] |
| 585 | ret |
| 586 | |
| 587 | $L$SEH_end_bn_mul4x_mont_gather5: |
| 588 | |
| 589 | |
| 590 | ALIGN 32 |
| 591 | mul4x_internal: |
| 592 | |
| 593 | shl r9,5 |
| 594 | movd xmm5,DWORD[56+rax] |
| 595 | lea rax,[$L$inc] |
| 596 | lea r13,[128+r9*1+rdx] |
| 597 | shr r9,5 |
| 598 | movdqa xmm0,XMMWORD[rax] |
| 599 | movdqa xmm1,XMMWORD[16+rax] |
| 600 | lea r10,[((88-112))+r9*1+rsp] |
| 601 | lea r12,[128+rdx] |
| 602 | |
| 603 | pshufd xmm5,xmm5,0 |
| 604 | movdqa xmm4,xmm1 |
| 605 | DB 0x67,0x67 |
| 606 | movdqa xmm2,xmm1 |
| 607 | paddd xmm1,xmm0 |
| 608 | pcmpeqd xmm0,xmm5 |
| 609 | DB 0x67 |
| 610 | movdqa xmm3,xmm4 |
| 611 | paddd xmm2,xmm1 |
| 612 | pcmpeqd xmm1,xmm5 |
| 613 | movdqa XMMWORD[112+r10],xmm0 |
| 614 | movdqa xmm0,xmm4 |
| 615 | |
| 616 | paddd xmm3,xmm2 |
| 617 | pcmpeqd xmm2,xmm5 |
| 618 | movdqa XMMWORD[128+r10],xmm1 |
| 619 | movdqa xmm1,xmm4 |
| 620 | |
| 621 | paddd xmm0,xmm3 |
| 622 | pcmpeqd xmm3,xmm5 |
| 623 | movdqa XMMWORD[144+r10],xmm2 |
| 624 | movdqa xmm2,xmm4 |
| 625 | |
| 626 | paddd xmm1,xmm0 |
| 627 | pcmpeqd xmm0,xmm5 |
| 628 | movdqa XMMWORD[160+r10],xmm3 |
| 629 | movdqa xmm3,xmm4 |
| 630 | paddd xmm2,xmm1 |
| 631 | pcmpeqd xmm1,xmm5 |
| 632 | movdqa XMMWORD[176+r10],xmm0 |
| 633 | movdqa xmm0,xmm4 |
| 634 | |
| 635 | paddd xmm3,xmm2 |
| 636 | pcmpeqd xmm2,xmm5 |
| 637 | movdqa XMMWORD[192+r10],xmm1 |
| 638 | movdqa xmm1,xmm4 |
| 639 | |
| 640 | paddd xmm0,xmm3 |
| 641 | pcmpeqd xmm3,xmm5 |
| 642 | movdqa XMMWORD[208+r10],xmm2 |
| 643 | movdqa xmm2,xmm4 |
| 644 | |
| 645 | paddd xmm1,xmm0 |
| 646 | pcmpeqd xmm0,xmm5 |
| 647 | movdqa XMMWORD[224+r10],xmm3 |
| 648 | movdqa xmm3,xmm4 |
| 649 | paddd xmm2,xmm1 |
| 650 | pcmpeqd xmm1,xmm5 |
| 651 | movdqa XMMWORD[240+r10],xmm0 |
| 652 | movdqa xmm0,xmm4 |
| 653 | |
| 654 | paddd xmm3,xmm2 |
| 655 | pcmpeqd xmm2,xmm5 |
| 656 | movdqa XMMWORD[256+r10],xmm1 |
| 657 | movdqa xmm1,xmm4 |
| 658 | |
| 659 | paddd xmm0,xmm3 |
| 660 | pcmpeqd xmm3,xmm5 |
| 661 | movdqa XMMWORD[272+r10],xmm2 |
| 662 | movdqa xmm2,xmm4 |
| 663 | |
| 664 | paddd xmm1,xmm0 |
| 665 | pcmpeqd xmm0,xmm5 |
| 666 | movdqa XMMWORD[288+r10],xmm3 |
| 667 | movdqa xmm3,xmm4 |
| 668 | paddd xmm2,xmm1 |
| 669 | pcmpeqd xmm1,xmm5 |
| 670 | movdqa XMMWORD[304+r10],xmm0 |
| 671 | |
| 672 | paddd xmm3,xmm2 |
| 673 | DB 0x67 |
| 674 | pcmpeqd xmm2,xmm5 |
| 675 | movdqa XMMWORD[320+r10],xmm1 |
| 676 | |
| 677 | pcmpeqd xmm3,xmm5 |
| 678 | movdqa XMMWORD[336+r10],xmm2 |
| 679 | pand xmm0,XMMWORD[64+r12] |
| 680 | |
| 681 | pand xmm1,XMMWORD[80+r12] |
| 682 | pand xmm2,XMMWORD[96+r12] |
| 683 | movdqa XMMWORD[352+r10],xmm3 |
| 684 | pand xmm3,XMMWORD[112+r12] |
| 685 | por xmm0,xmm2 |
| 686 | por xmm1,xmm3 |
| 687 | movdqa xmm4,XMMWORD[((-128))+r12] |
| 688 | movdqa xmm5,XMMWORD[((-112))+r12] |
| 689 | movdqa xmm2,XMMWORD[((-96))+r12] |
| 690 | pand xmm4,XMMWORD[112+r10] |
| 691 | movdqa xmm3,XMMWORD[((-80))+r12] |
| 692 | pand xmm5,XMMWORD[128+r10] |
| 693 | por xmm0,xmm4 |
| 694 | pand xmm2,XMMWORD[144+r10] |
| 695 | por xmm1,xmm5 |
| 696 | pand xmm3,XMMWORD[160+r10] |
| 697 | por xmm0,xmm2 |
| 698 | por xmm1,xmm3 |
| 699 | movdqa xmm4,XMMWORD[((-64))+r12] |
| 700 | movdqa xmm5,XMMWORD[((-48))+r12] |
| 701 | movdqa xmm2,XMMWORD[((-32))+r12] |
| 702 | pand xmm4,XMMWORD[176+r10] |
| 703 | movdqa xmm3,XMMWORD[((-16))+r12] |
| 704 | pand xmm5,XMMWORD[192+r10] |
| 705 | por xmm0,xmm4 |
| 706 | pand xmm2,XMMWORD[208+r10] |
| 707 | por xmm1,xmm5 |
| 708 | pand xmm3,XMMWORD[224+r10] |
| 709 | por xmm0,xmm2 |
| 710 | por xmm1,xmm3 |
| 711 | movdqa xmm4,XMMWORD[r12] |
| 712 | movdqa xmm5,XMMWORD[16+r12] |
| 713 | movdqa xmm2,XMMWORD[32+r12] |
| 714 | pand xmm4,XMMWORD[240+r10] |
| 715 | movdqa xmm3,XMMWORD[48+r12] |
| 716 | pand xmm5,XMMWORD[256+r10] |
| 717 | por xmm0,xmm4 |
| 718 | pand xmm2,XMMWORD[272+r10] |
| 719 | por xmm1,xmm5 |
| 720 | pand xmm3,XMMWORD[288+r10] |
| 721 | por xmm0,xmm2 |
| 722 | por xmm1,xmm3 |
| 723 | por xmm0,xmm1 |
| 724 | |
| 725 | pshufd xmm1,xmm0,0x4e |
| 726 | por xmm0,xmm1 |
| 727 | lea r12,[256+r12] |
| 728 | DB 102,72,15,126,195 |
| 729 | |
| 730 | mov QWORD[((16+8))+rsp],r13 |
| 731 | mov QWORD[((56+8))+rsp],rdi |
| 732 | |
| 733 | mov r8,QWORD[r8] |
| 734 | mov rax,QWORD[rsi] |
| 735 | lea rsi,[r9*1+rsi] |
| 736 | neg r9 |
| 737 | |
| 738 | mov rbp,r8 |
| 739 | mul rbx |
| 740 | mov r10,rax |
| 741 | mov rax,QWORD[rcx] |
| 742 | |
| 743 | imul rbp,r10 |
| 744 | lea r14,[((64+8))+rsp] |
| 745 | mov r11,rdx |
| 746 | |
| 747 | mul rbp |
| 748 | add r10,rax |
| 749 | mov rax,QWORD[8+r9*1+rsi] |
| 750 | adc rdx,0 |
| 751 | mov rdi,rdx |
| 752 | |
| 753 | mul rbx |
| 754 | add r11,rax |
| 755 | mov rax,QWORD[8+rcx] |
| 756 | adc rdx,0 |
| 757 | mov r10,rdx |
| 758 | |
| 759 | mul rbp |
| 760 | add rdi,rax |
| 761 | mov rax,QWORD[16+r9*1+rsi] |
| 762 | adc rdx,0 |
| 763 | add rdi,r11 |
| 764 | lea r15,[32+r9] |
| 765 | lea rcx,[32+rcx] |
| 766 | adc rdx,0 |
| 767 | mov QWORD[r14],rdi |
| 768 | mov r13,rdx |
| 769 | jmp NEAR $L$1st4x |
| 770 | |
| 771 | ALIGN 32 |
| 772 | $L$1st4x: |
| 773 | mul rbx |
| 774 | add r10,rax |
| 775 | mov rax,QWORD[((-16))+rcx] |
| 776 | lea r14,[32+r14] |
| 777 | adc rdx,0 |
| 778 | mov r11,rdx |
| 779 | |
| 780 | mul rbp |
| 781 | add r13,rax |
| 782 | mov rax,QWORD[((-8))+r15*1+rsi] |
| 783 | adc rdx,0 |
| 784 | add r13,r10 |
| 785 | adc rdx,0 |
| 786 | mov QWORD[((-24))+r14],r13 |
| 787 | mov rdi,rdx |
| 788 | |
| 789 | mul rbx |
| 790 | add r11,rax |
| 791 | mov rax,QWORD[((-8))+rcx] |
| 792 | adc rdx,0 |
| 793 | mov r10,rdx |
| 794 | |
| 795 | mul rbp |
| 796 | add rdi,rax |
| 797 | mov rax,QWORD[r15*1+rsi] |
| 798 | adc rdx,0 |
| 799 | add rdi,r11 |
| 800 | adc rdx,0 |
| 801 | mov QWORD[((-16))+r14],rdi |
| 802 | mov r13,rdx |
| 803 | |
| 804 | mul rbx |
| 805 | add r10,rax |
| 806 | mov rax,QWORD[rcx] |
| 807 | adc rdx,0 |
| 808 | mov r11,rdx |
| 809 | |
| 810 | mul rbp |
| 811 | add r13,rax |
| 812 | mov rax,QWORD[8+r15*1+rsi] |
| 813 | adc rdx,0 |
| 814 | add r13,r10 |
| 815 | adc rdx,0 |
| 816 | mov QWORD[((-8))+r14],r13 |
| 817 | mov rdi,rdx |
| 818 | |
| 819 | mul rbx |
| 820 | add r11,rax |
| 821 | mov rax,QWORD[8+rcx] |
| 822 | adc rdx,0 |
| 823 | mov r10,rdx |
| 824 | |
| 825 | mul rbp |
| 826 | add rdi,rax |
| 827 | mov rax,QWORD[16+r15*1+rsi] |
| 828 | adc rdx,0 |
| 829 | add rdi,r11 |
| 830 | lea rcx,[32+rcx] |
| 831 | adc rdx,0 |
| 832 | mov QWORD[r14],rdi |
| 833 | mov r13,rdx |
| 834 | |
| 835 | add r15,32 |
| 836 | jnz NEAR $L$1st4x |
| 837 | |
| 838 | mul rbx |
| 839 | add r10,rax |
| 840 | mov rax,QWORD[((-16))+rcx] |
| 841 | lea r14,[32+r14] |
| 842 | adc rdx,0 |
| 843 | mov r11,rdx |
| 844 | |
| 845 | mul rbp |
| 846 | add r13,rax |
| 847 | mov rax,QWORD[((-8))+rsi] |
| 848 | adc rdx,0 |
| 849 | add r13,r10 |
| 850 | adc rdx,0 |
| 851 | mov QWORD[((-24))+r14],r13 |
| 852 | mov rdi,rdx |
| 853 | |
| 854 | mul rbx |
| 855 | add r11,rax |
| 856 | mov rax,QWORD[((-8))+rcx] |
| 857 | adc rdx,0 |
| 858 | mov r10,rdx |
| 859 | |
| 860 | mul rbp |
| 861 | add rdi,rax |
| 862 | mov rax,QWORD[r9*1+rsi] |
| 863 | adc rdx,0 |
| 864 | add rdi,r11 |
| 865 | adc rdx,0 |
| 866 | mov QWORD[((-16))+r14],rdi |
| 867 | mov r13,rdx |
| 868 | |
| 869 | lea rcx,[r9*1+rcx] |
| 870 | |
| 871 | xor rdi,rdi |
| 872 | add r13,r10 |
| 873 | adc rdi,0 |
| 874 | mov QWORD[((-8))+r14],r13 |
| 875 | |
| 876 | jmp NEAR $L$outer4x |
| 877 | |
| 878 | ALIGN 32 |
| 879 | $L$outer4x: |
| 880 | lea rdx,[((16+128))+r14] |
| 881 | pxor xmm4,xmm4 |
| 882 | pxor xmm5,xmm5 |
| 883 | movdqa xmm0,XMMWORD[((-128))+r12] |
| 884 | movdqa xmm1,XMMWORD[((-112))+r12] |
| 885 | movdqa xmm2,XMMWORD[((-96))+r12] |
| 886 | movdqa xmm3,XMMWORD[((-80))+r12] |
| 887 | pand xmm0,XMMWORD[((-128))+rdx] |
| 888 | pand xmm1,XMMWORD[((-112))+rdx] |
| 889 | por xmm4,xmm0 |
| 890 | pand xmm2,XMMWORD[((-96))+rdx] |
| 891 | por xmm5,xmm1 |
| 892 | pand xmm3,XMMWORD[((-80))+rdx] |
| 893 | por xmm4,xmm2 |
| 894 | por xmm5,xmm3 |
| 895 | movdqa xmm0,XMMWORD[((-64))+r12] |
| 896 | movdqa xmm1,XMMWORD[((-48))+r12] |
| 897 | movdqa xmm2,XMMWORD[((-32))+r12] |
| 898 | movdqa xmm3,XMMWORD[((-16))+r12] |
| 899 | pand xmm0,XMMWORD[((-64))+rdx] |
| 900 | pand xmm1,XMMWORD[((-48))+rdx] |
| 901 | por xmm4,xmm0 |
| 902 | pand xmm2,XMMWORD[((-32))+rdx] |
| 903 | por xmm5,xmm1 |
| 904 | pand xmm3,XMMWORD[((-16))+rdx] |
| 905 | por xmm4,xmm2 |
| 906 | por xmm5,xmm3 |
| 907 | movdqa xmm0,XMMWORD[r12] |
| 908 | movdqa xmm1,XMMWORD[16+r12] |
| 909 | movdqa xmm2,XMMWORD[32+r12] |
| 910 | movdqa xmm3,XMMWORD[48+r12] |
| 911 | pand xmm0,XMMWORD[rdx] |
| 912 | pand xmm1,XMMWORD[16+rdx] |
| 913 | por xmm4,xmm0 |
| 914 | pand xmm2,XMMWORD[32+rdx] |
| 915 | por xmm5,xmm1 |
| 916 | pand xmm3,XMMWORD[48+rdx] |
| 917 | por xmm4,xmm2 |
| 918 | por xmm5,xmm3 |
| 919 | movdqa xmm0,XMMWORD[64+r12] |
| 920 | movdqa xmm1,XMMWORD[80+r12] |
| 921 | movdqa xmm2,XMMWORD[96+r12] |
| 922 | movdqa xmm3,XMMWORD[112+r12] |
| 923 | pand xmm0,XMMWORD[64+rdx] |
| 924 | pand xmm1,XMMWORD[80+rdx] |
| 925 | por xmm4,xmm0 |
| 926 | pand xmm2,XMMWORD[96+rdx] |
| 927 | por xmm5,xmm1 |
| 928 | pand xmm3,XMMWORD[112+rdx] |
| 929 | por xmm4,xmm2 |
| 930 | por xmm5,xmm3 |
| 931 | por xmm4,xmm5 |
| 932 | |
| 933 | pshufd xmm0,xmm4,0x4e |
| 934 | por xmm0,xmm4 |
| 935 | lea r12,[256+r12] |
| 936 | DB 102,72,15,126,195 |
| 937 | |
| 938 | mov r10,QWORD[r9*1+r14] |
| 939 | mov rbp,r8 |
| 940 | mul rbx |
| 941 | add r10,rax |
| 942 | mov rax,QWORD[rcx] |
| 943 | adc rdx,0 |
| 944 | |
| 945 | imul rbp,r10 |
| 946 | mov r11,rdx |
| 947 | mov QWORD[r14],rdi |
| 948 | |
| 949 | lea r14,[r9*1+r14] |
| 950 | |
| 951 | mul rbp |
| 952 | add r10,rax |
| 953 | mov rax,QWORD[8+r9*1+rsi] |
| 954 | adc rdx,0 |
| 955 | mov rdi,rdx |
| 956 | |
| 957 | mul rbx |
| 958 | add r11,rax |
| 959 | mov rax,QWORD[8+rcx] |
| 960 | adc rdx,0 |
| 961 | add r11,QWORD[8+r14] |
| 962 | adc rdx,0 |
| 963 | mov r10,rdx |
| 964 | |
| 965 | mul rbp |
| 966 | add rdi,rax |
| 967 | mov rax,QWORD[16+r9*1+rsi] |
| 968 | adc rdx,0 |
| 969 | add rdi,r11 |
| 970 | lea r15,[32+r9] |
| 971 | lea rcx,[32+rcx] |
| 972 | adc rdx,0 |
| 973 | mov r13,rdx |
| 974 | jmp NEAR $L$inner4x |
| 975 | |
| 976 | ALIGN 32 |
| 977 | $L$inner4x: |
| 978 | mul rbx |
| 979 | add r10,rax |
| 980 | mov rax,QWORD[((-16))+rcx] |
| 981 | adc rdx,0 |
| 982 | add r10,QWORD[16+r14] |
| 983 | lea r14,[32+r14] |
| 984 | adc rdx,0 |
| 985 | mov r11,rdx |
| 986 | |
| 987 | mul rbp |
| 988 | add r13,rax |
| 989 | mov rax,QWORD[((-8))+r15*1+rsi] |
| 990 | adc rdx,0 |
| 991 | add r13,r10 |
| 992 | adc rdx,0 |
| 993 | mov QWORD[((-32))+r14],rdi |
| 994 | mov rdi,rdx |
| 995 | |
| 996 | mul rbx |
| 997 | add r11,rax |
| 998 | mov rax,QWORD[((-8))+rcx] |
| 999 | adc rdx,0 |
| 1000 | add r11,QWORD[((-8))+r14] |
| 1001 | adc rdx,0 |
| 1002 | mov r10,rdx |
| 1003 | |
| 1004 | mul rbp |
| 1005 | add rdi,rax |
| 1006 | mov rax,QWORD[r15*1+rsi] |
| 1007 | adc rdx,0 |
| 1008 | add rdi,r11 |
| 1009 | adc rdx,0 |
| 1010 | mov QWORD[((-24))+r14],r13 |
| 1011 | mov r13,rdx |
| 1012 | |
| 1013 | mul rbx |
| 1014 | add r10,rax |
| 1015 | mov rax,QWORD[rcx] |
| 1016 | adc rdx,0 |
| 1017 | add r10,QWORD[r14] |
| 1018 | adc rdx,0 |
| 1019 | mov r11,rdx |
| 1020 | |
| 1021 | mul rbp |
| 1022 | add r13,rax |
| 1023 | mov rax,QWORD[8+r15*1+rsi] |
| 1024 | adc rdx,0 |
| 1025 | add r13,r10 |
| 1026 | adc rdx,0 |
| 1027 | mov QWORD[((-16))+r14],rdi |
| 1028 | mov rdi,rdx |
| 1029 | |
| 1030 | mul rbx |
| 1031 | add r11,rax |
| 1032 | mov rax,QWORD[8+rcx] |
| 1033 | adc rdx,0 |
| 1034 | add r11,QWORD[8+r14] |
| 1035 | adc rdx,0 |
| 1036 | mov r10,rdx |
| 1037 | |
| 1038 | mul rbp |
| 1039 | add rdi,rax |
| 1040 | mov rax,QWORD[16+r15*1+rsi] |
| 1041 | adc rdx,0 |
| 1042 | add rdi,r11 |
| 1043 | lea rcx,[32+rcx] |
| 1044 | adc rdx,0 |
| 1045 | mov QWORD[((-8))+r14],r13 |
| 1046 | mov r13,rdx |
| 1047 | |
| 1048 | add r15,32 |
| 1049 | jnz NEAR $L$inner4x |
| 1050 | |
| 1051 | mul rbx |
| 1052 | add r10,rax |
| 1053 | mov rax,QWORD[((-16))+rcx] |
| 1054 | adc rdx,0 |
| 1055 | add r10,QWORD[16+r14] |
| 1056 | lea r14,[32+r14] |
| 1057 | adc rdx,0 |
| 1058 | mov r11,rdx |
| 1059 | |
| 1060 | mul rbp |
| 1061 | add r13,rax |
| 1062 | mov rax,QWORD[((-8))+rsi] |
| 1063 | adc rdx,0 |
| 1064 | add r13,r10 |
| 1065 | adc rdx,0 |
| 1066 | mov QWORD[((-32))+r14],rdi |
| 1067 | mov rdi,rdx |
| 1068 | |
| 1069 | mul rbx |
| 1070 | add r11,rax |
| 1071 | mov rax,rbp |
| 1072 | mov rbp,QWORD[((-8))+rcx] |
| 1073 | adc rdx,0 |
| 1074 | add r11,QWORD[((-8))+r14] |
| 1075 | adc rdx,0 |
| 1076 | mov r10,rdx |
| 1077 | |
| 1078 | mul rbp |
| 1079 | add rdi,rax |
| 1080 | mov rax,QWORD[r9*1+rsi] |
| 1081 | adc rdx,0 |
| 1082 | add rdi,r11 |
| 1083 | adc rdx,0 |
| 1084 | mov QWORD[((-24))+r14],r13 |
| 1085 | mov r13,rdx |
| 1086 | |
| 1087 | mov QWORD[((-16))+r14],rdi |
| 1088 | lea rcx,[r9*1+rcx] |
| 1089 | |
| 1090 | xor rdi,rdi |
| 1091 | add r13,r10 |
| 1092 | adc rdi,0 |
| 1093 | add r13,QWORD[r14] |
| 1094 | adc rdi,0 |
| 1095 | mov QWORD[((-8))+r14],r13 |
| 1096 | |
| 1097 | cmp r12,QWORD[((16+8))+rsp] |
| 1098 | jb NEAR $L$outer4x |
| 1099 | xor rax,rax |
| 1100 | sub rbp,r13 |
| 1101 | adc r15,r15 |
| 1102 | or rdi,r15 |
| 1103 | sub rax,rdi |
| 1104 | lea rbx,[r9*1+r14] |
| 1105 | mov r12,QWORD[rcx] |
| 1106 | lea rbp,[rcx] |
| 1107 | mov rcx,r9 |
| 1108 | sar rcx,3+2 |
| 1109 | mov rdi,QWORD[((56+8))+rsp] |
| 1110 | dec r12 |
| 1111 | xor r10,r10 |
| 1112 | mov r13,QWORD[8+rbp] |
| 1113 | mov r14,QWORD[16+rbp] |
| 1114 | mov r15,QWORD[24+rbp] |
| 1115 | jmp NEAR $L$sqr4x_sub_entry |
| 1116 | |
| 1117 | |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 1118 | global bn_power5_nohw |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 1119 | |
| 1120 | ALIGN 32 |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 1121 | bn_power5_nohw: |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 1122 | mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 1123 | mov QWORD[16+rsp],rsi |
| 1124 | mov rax,rsp |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 1125 | $L$SEH_begin_bn_power5_nohw: |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 1126 | mov rdi,rcx |
| 1127 | mov rsi,rdx |
| 1128 | mov rdx,r8 |
| 1129 | mov rcx,r9 |
| 1130 | mov r8,QWORD[40+rsp] |
| 1131 | mov r9,QWORD[48+rsp] |
| 1132 | |
| 1133 | |
| 1134 | |
| 1135 | _CET_ENDBR |
| 1136 | mov rax,rsp |
| 1137 | |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 1138 | push rbx |
| 1139 | |
| 1140 | push rbp |
| 1141 | |
| 1142 | push r12 |
| 1143 | |
| 1144 | push r13 |
| 1145 | |
| 1146 | push r14 |
| 1147 | |
| 1148 | push r15 |
| 1149 | |
| 1150 | $L$power5_prologue: |
| 1151 | |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 1152 | |
| 1153 | |
| 1154 | |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 1155 | shl r9d,3 |
| 1156 | lea r10d,[r9*2+r9] |
| 1157 | neg r9 |
| 1158 | mov r8,QWORD[r8] |
| 1159 | |
| 1160 | |
| 1161 | |
| 1162 | |
| 1163 | |
| 1164 | |
| 1165 | |
| 1166 | |
| 1167 | lea r11,[((-320))+r9*2+rsp] |
| 1168 | mov rbp,rsp |
| 1169 | sub r11,rdi |
| 1170 | and r11,4095 |
| 1171 | cmp r10,r11 |
| 1172 | jb NEAR $L$pwr_sp_alt |
| 1173 | sub rbp,r11 |
| 1174 | lea rbp,[((-320))+r9*2+rbp] |
| 1175 | jmp NEAR $L$pwr_sp_done |
| 1176 | |
| 1177 | ALIGN 32 |
| 1178 | $L$pwr_sp_alt: |
| 1179 | lea r10,[((4096-320))+r9*2] |
| 1180 | lea rbp,[((-320))+r9*2+rbp] |
| 1181 | sub r11,r10 |
| 1182 | mov r10,0 |
| 1183 | cmovc r11,r10 |
| 1184 | sub rbp,r11 |
| 1185 | $L$pwr_sp_done: |
| 1186 | and rbp,-64 |
| 1187 | mov r11,rsp |
| 1188 | sub r11,rbp |
| 1189 | and r11,-4096 |
| 1190 | lea rsp,[rbp*1+r11] |
| 1191 | mov r10,QWORD[rsp] |
| 1192 | cmp rsp,rbp |
| 1193 | ja NEAR $L$pwr_page_walk |
| 1194 | jmp NEAR $L$pwr_page_walk_done |
| 1195 | |
| 1196 | $L$pwr_page_walk: |
| 1197 | lea rsp,[((-4096))+rsp] |
| 1198 | mov r10,QWORD[rsp] |
| 1199 | cmp rsp,rbp |
| 1200 | ja NEAR $L$pwr_page_walk |
| 1201 | $L$pwr_page_walk_done: |
| 1202 | |
| 1203 | mov r10,r9 |
| 1204 | neg r9 |
| 1205 | |
| 1206 | |
| 1207 | |
| 1208 | |
| 1209 | |
| 1210 | |
| 1211 | |
| 1212 | |
| 1213 | |
| 1214 | |
| 1215 | mov QWORD[32+rsp],r8 |
| 1216 | mov QWORD[40+rsp],rax |
| 1217 | |
| 1218 | $L$power5_body: |
| 1219 | DB 102,72,15,110,207 |
| 1220 | DB 102,72,15,110,209 |
| 1221 | DB 102,73,15,110,218 |
| 1222 | DB 102,72,15,110,226 |
| 1223 | |
| 1224 | call __bn_sqr8x_internal |
| 1225 | call __bn_post4x_internal |
| 1226 | call __bn_sqr8x_internal |
| 1227 | call __bn_post4x_internal |
| 1228 | call __bn_sqr8x_internal |
| 1229 | call __bn_post4x_internal |
| 1230 | call __bn_sqr8x_internal |
| 1231 | call __bn_post4x_internal |
| 1232 | call __bn_sqr8x_internal |
| 1233 | call __bn_post4x_internal |
| 1234 | |
| 1235 | DB 102,72,15,126,209 |
| 1236 | DB 102,72,15,126,226 |
| 1237 | mov rdi,rsi |
| 1238 | mov rax,QWORD[40+rsp] |
| 1239 | lea r8,[32+rsp] |
| 1240 | |
| 1241 | call mul4x_internal |
| 1242 | |
| 1243 | mov rsi,QWORD[40+rsp] |
| 1244 | |
| 1245 | mov rax,1 |
| 1246 | mov r15,QWORD[((-48))+rsi] |
| 1247 | |
| 1248 | mov r14,QWORD[((-40))+rsi] |
| 1249 | |
| 1250 | mov r13,QWORD[((-32))+rsi] |
| 1251 | |
| 1252 | mov r12,QWORD[((-24))+rsi] |
| 1253 | |
| 1254 | mov rbp,QWORD[((-16))+rsi] |
| 1255 | |
| 1256 | mov rbx,QWORD[((-8))+rsi] |
| 1257 | |
| 1258 | lea rsp,[rsi] |
| 1259 | |
| 1260 | $L$power5_epilogue: |
| 1261 | mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 1262 | mov rsi,QWORD[16+rsp] |
| 1263 | ret |
| 1264 | |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 1265 | $L$SEH_end_bn_power5_nohw: |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 1266 | |
| 1267 | global bn_sqr8x_internal |
| 1268 | |
| 1269 | |
| 1270 | ALIGN 32 |
| 1271 | bn_sqr8x_internal: |
| 1272 | __bn_sqr8x_internal: |
| 1273 | |
| 1274 | _CET_ENDBR |
| 1275 | |
| 1276 | |
| 1277 | |
| 1278 | |
| 1279 | |
| 1280 | |
| 1281 | |
| 1282 | |
| 1283 | |
| 1284 | |
| 1285 | |
| 1286 | |
| 1287 | |
| 1288 | |
| 1289 | |
| 1290 | |
| 1291 | |
| 1292 | |
| 1293 | |
| 1294 | |
| 1295 | |
| 1296 | |
| 1297 | |
| 1298 | |
| 1299 | |
| 1300 | |
| 1301 | |
| 1302 | |
| 1303 | |
| 1304 | |
| 1305 | |
| 1306 | |
| 1307 | |
| 1308 | |
| 1309 | |
| 1310 | |
| 1311 | |
| 1312 | |
| 1313 | |
| 1314 | |
| 1315 | |
| 1316 | |
| 1317 | |
| 1318 | |
| 1319 | |
| 1320 | |
| 1321 | |
| 1322 | |
| 1323 | |
| 1324 | |
| 1325 | |
| 1326 | |
| 1327 | |
| 1328 | |
| 1329 | |
| 1330 | |
| 1331 | |
| 1332 | |
| 1333 | |
| 1334 | |
| 1335 | |
| 1336 | |
| 1337 | |
| 1338 | |
| 1339 | |
| 1340 | |
| 1341 | |
| 1342 | |
| 1343 | |
| 1344 | |
| 1345 | |
| 1346 | |
| 1347 | |
| 1348 | lea rbp,[32+r10] |
| 1349 | lea rsi,[r9*1+rsi] |
| 1350 | |
| 1351 | mov rcx,r9 |
| 1352 | |
| 1353 | |
| 1354 | mov r14,QWORD[((-32))+rbp*1+rsi] |
| 1355 | lea rdi,[((48+8))+r9*2+rsp] |
| 1356 | mov rax,QWORD[((-24))+rbp*1+rsi] |
| 1357 | lea rdi,[((-32))+rbp*1+rdi] |
| 1358 | mov rbx,QWORD[((-16))+rbp*1+rsi] |
| 1359 | mov r15,rax |
| 1360 | |
| 1361 | mul r14 |
| 1362 | mov r10,rax |
| 1363 | mov rax,rbx |
| 1364 | mov r11,rdx |
| 1365 | mov QWORD[((-24))+rbp*1+rdi],r10 |
| 1366 | |
| 1367 | mul r14 |
| 1368 | add r11,rax |
| 1369 | mov rax,rbx |
| 1370 | adc rdx,0 |
| 1371 | mov QWORD[((-16))+rbp*1+rdi],r11 |
| 1372 | mov r10,rdx |
| 1373 | |
| 1374 | |
| 1375 | mov rbx,QWORD[((-8))+rbp*1+rsi] |
| 1376 | mul r15 |
| 1377 | mov r12,rax |
| 1378 | mov rax,rbx |
| 1379 | mov r13,rdx |
| 1380 | |
| 1381 | lea rcx,[rbp] |
| 1382 | mul r14 |
| 1383 | add r10,rax |
| 1384 | mov rax,rbx |
| 1385 | mov r11,rdx |
| 1386 | adc r11,0 |
| 1387 | add r10,r12 |
| 1388 | adc r11,0 |
| 1389 | mov QWORD[((-8))+rcx*1+rdi],r10 |
| 1390 | jmp NEAR $L$sqr4x_1st |
| 1391 | |
| 1392 | ALIGN 32 |
| 1393 | $L$sqr4x_1st: |
| 1394 | mov rbx,QWORD[rcx*1+rsi] |
| 1395 | mul r15 |
| 1396 | add r13,rax |
| 1397 | mov rax,rbx |
| 1398 | mov r12,rdx |
| 1399 | adc r12,0 |
| 1400 | |
| 1401 | mul r14 |
| 1402 | add r11,rax |
| 1403 | mov rax,rbx |
| 1404 | mov rbx,QWORD[8+rcx*1+rsi] |
| 1405 | mov r10,rdx |
| 1406 | adc r10,0 |
| 1407 | add r11,r13 |
| 1408 | adc r10,0 |
| 1409 | |
| 1410 | |
| 1411 | mul r15 |
| 1412 | add r12,rax |
| 1413 | mov rax,rbx |
| 1414 | mov QWORD[rcx*1+rdi],r11 |
| 1415 | mov r13,rdx |
| 1416 | adc r13,0 |
| 1417 | |
| 1418 | mul r14 |
| 1419 | add r10,rax |
| 1420 | mov rax,rbx |
| 1421 | mov rbx,QWORD[16+rcx*1+rsi] |
| 1422 | mov r11,rdx |
| 1423 | adc r11,0 |
| 1424 | add r10,r12 |
| 1425 | adc r11,0 |
| 1426 | |
| 1427 | mul r15 |
| 1428 | add r13,rax |
| 1429 | mov rax,rbx |
| 1430 | mov QWORD[8+rcx*1+rdi],r10 |
| 1431 | mov r12,rdx |
| 1432 | adc r12,0 |
| 1433 | |
| 1434 | mul r14 |
| 1435 | add r11,rax |
| 1436 | mov rax,rbx |
| 1437 | mov rbx,QWORD[24+rcx*1+rsi] |
| 1438 | mov r10,rdx |
| 1439 | adc r10,0 |
| 1440 | add r11,r13 |
| 1441 | adc r10,0 |
| 1442 | |
| 1443 | |
| 1444 | mul r15 |
| 1445 | add r12,rax |
| 1446 | mov rax,rbx |
| 1447 | mov QWORD[16+rcx*1+rdi],r11 |
| 1448 | mov r13,rdx |
| 1449 | adc r13,0 |
| 1450 | lea rcx,[32+rcx] |
| 1451 | |
| 1452 | mul r14 |
| 1453 | add r10,rax |
| 1454 | mov rax,rbx |
| 1455 | mov r11,rdx |
| 1456 | adc r11,0 |
| 1457 | add r10,r12 |
| 1458 | adc r11,0 |
| 1459 | mov QWORD[((-8))+rcx*1+rdi],r10 |
| 1460 | |
| 1461 | cmp rcx,0 |
| 1462 | jne NEAR $L$sqr4x_1st |
| 1463 | |
| 1464 | mul r15 |
| 1465 | add r13,rax |
| 1466 | lea rbp,[16+rbp] |
| 1467 | adc rdx,0 |
| 1468 | add r13,r11 |
| 1469 | adc rdx,0 |
| 1470 | |
| 1471 | mov QWORD[rdi],r13 |
| 1472 | mov r12,rdx |
| 1473 | mov QWORD[8+rdi],rdx |
| 1474 | jmp NEAR $L$sqr4x_outer |
| 1475 | |
| 1476 | ALIGN 32 |
| 1477 | $L$sqr4x_outer: |
| 1478 | mov r14,QWORD[((-32))+rbp*1+rsi] |
| 1479 | lea rdi,[((48+8))+r9*2+rsp] |
| 1480 | mov rax,QWORD[((-24))+rbp*1+rsi] |
| 1481 | lea rdi,[((-32))+rbp*1+rdi] |
| 1482 | mov rbx,QWORD[((-16))+rbp*1+rsi] |
| 1483 | mov r15,rax |
| 1484 | |
| 1485 | mul r14 |
| 1486 | mov r10,QWORD[((-24))+rbp*1+rdi] |
| 1487 | add r10,rax |
| 1488 | mov rax,rbx |
| 1489 | adc rdx,0 |
| 1490 | mov QWORD[((-24))+rbp*1+rdi],r10 |
| 1491 | mov r11,rdx |
| 1492 | |
| 1493 | mul r14 |
| 1494 | add r11,rax |
| 1495 | mov rax,rbx |
| 1496 | adc rdx,0 |
| 1497 | add r11,QWORD[((-16))+rbp*1+rdi] |
| 1498 | mov r10,rdx |
| 1499 | adc r10,0 |
| 1500 | mov QWORD[((-16))+rbp*1+rdi],r11 |
| 1501 | |
| 1502 | xor r12,r12 |
| 1503 | |
| 1504 | mov rbx,QWORD[((-8))+rbp*1+rsi] |
| 1505 | mul r15 |
| 1506 | add r12,rax |
| 1507 | mov rax,rbx |
| 1508 | adc rdx,0 |
| 1509 | add r12,QWORD[((-8))+rbp*1+rdi] |
| 1510 | mov r13,rdx |
| 1511 | adc r13,0 |
| 1512 | |
| 1513 | mul r14 |
| 1514 | add r10,rax |
| 1515 | mov rax,rbx |
| 1516 | adc rdx,0 |
| 1517 | add r10,r12 |
| 1518 | mov r11,rdx |
| 1519 | adc r11,0 |
| 1520 | mov QWORD[((-8))+rbp*1+rdi],r10 |
| 1521 | |
| 1522 | lea rcx,[rbp] |
| 1523 | jmp NEAR $L$sqr4x_inner |
| 1524 | |
| 1525 | ALIGN 32 |
| 1526 | $L$sqr4x_inner: |
| 1527 | mov rbx,QWORD[rcx*1+rsi] |
| 1528 | mul r15 |
| 1529 | add r13,rax |
| 1530 | mov rax,rbx |
| 1531 | mov r12,rdx |
| 1532 | adc r12,0 |
| 1533 | add r13,QWORD[rcx*1+rdi] |
| 1534 | adc r12,0 |
| 1535 | |
| 1536 | DB 0x67 |
| 1537 | mul r14 |
| 1538 | add r11,rax |
| 1539 | mov rax,rbx |
| 1540 | mov rbx,QWORD[8+rcx*1+rsi] |
| 1541 | mov r10,rdx |
| 1542 | adc r10,0 |
| 1543 | add r11,r13 |
| 1544 | adc r10,0 |
| 1545 | |
| 1546 | mul r15 |
| 1547 | add r12,rax |
| 1548 | mov QWORD[rcx*1+rdi],r11 |
| 1549 | mov rax,rbx |
| 1550 | mov r13,rdx |
| 1551 | adc r13,0 |
| 1552 | add r12,QWORD[8+rcx*1+rdi] |
| 1553 | lea rcx,[16+rcx] |
| 1554 | adc r13,0 |
| 1555 | |
| 1556 | mul r14 |
| 1557 | add r10,rax |
| 1558 | mov rax,rbx |
| 1559 | adc rdx,0 |
| 1560 | add r10,r12 |
| 1561 | mov r11,rdx |
| 1562 | adc r11,0 |
| 1563 | mov QWORD[((-8))+rcx*1+rdi],r10 |
| 1564 | |
| 1565 | cmp rcx,0 |
| 1566 | jne NEAR $L$sqr4x_inner |
| 1567 | |
| 1568 | DB 0x67 |
| 1569 | mul r15 |
| 1570 | add r13,rax |
| 1571 | adc rdx,0 |
| 1572 | add r13,r11 |
| 1573 | adc rdx,0 |
| 1574 | |
| 1575 | mov QWORD[rdi],r13 |
| 1576 | mov r12,rdx |
| 1577 | mov QWORD[8+rdi],rdx |
| 1578 | |
| 1579 | add rbp,16 |
| 1580 | jnz NEAR $L$sqr4x_outer |
| 1581 | |
| 1582 | |
| 1583 | mov r14,QWORD[((-32))+rsi] |
| 1584 | lea rdi,[((48+8))+r9*2+rsp] |
| 1585 | mov rax,QWORD[((-24))+rsi] |
| 1586 | lea rdi,[((-32))+rbp*1+rdi] |
| 1587 | mov rbx,QWORD[((-16))+rsi] |
| 1588 | mov r15,rax |
| 1589 | |
| 1590 | mul r14 |
| 1591 | add r10,rax |
| 1592 | mov rax,rbx |
| 1593 | mov r11,rdx |
| 1594 | adc r11,0 |
| 1595 | |
| 1596 | mul r14 |
| 1597 | add r11,rax |
| 1598 | mov rax,rbx |
| 1599 | mov QWORD[((-24))+rdi],r10 |
| 1600 | mov r10,rdx |
| 1601 | adc r10,0 |
| 1602 | add r11,r13 |
| 1603 | mov rbx,QWORD[((-8))+rsi] |
| 1604 | adc r10,0 |
| 1605 | |
| 1606 | mul r15 |
| 1607 | add r12,rax |
| 1608 | mov rax,rbx |
| 1609 | mov QWORD[((-16))+rdi],r11 |
| 1610 | mov r13,rdx |
| 1611 | adc r13,0 |
| 1612 | |
| 1613 | mul r14 |
| 1614 | add r10,rax |
| 1615 | mov rax,rbx |
| 1616 | mov r11,rdx |
| 1617 | adc r11,0 |
| 1618 | add r10,r12 |
| 1619 | adc r11,0 |
| 1620 | mov QWORD[((-8))+rdi],r10 |
| 1621 | |
| 1622 | mul r15 |
| 1623 | add r13,rax |
| 1624 | mov rax,QWORD[((-16))+rsi] |
| 1625 | adc rdx,0 |
| 1626 | add r13,r11 |
| 1627 | adc rdx,0 |
| 1628 | |
| 1629 | mov QWORD[rdi],r13 |
| 1630 | mov r12,rdx |
| 1631 | mov QWORD[8+rdi],rdx |
| 1632 | |
| 1633 | mul rbx |
| 1634 | add rbp,16 |
| 1635 | xor r14,r14 |
| 1636 | sub rbp,r9 |
| 1637 | xor r15,r15 |
| 1638 | |
| 1639 | add rax,r12 |
| 1640 | adc rdx,0 |
| 1641 | mov QWORD[8+rdi],rax |
| 1642 | mov QWORD[16+rdi],rdx |
| 1643 | mov QWORD[24+rdi],r15 |
| 1644 | |
| 1645 | mov rax,QWORD[((-16))+rbp*1+rsi] |
| 1646 | lea rdi,[((48+8))+rsp] |
| 1647 | xor r10,r10 |
| 1648 | mov r11,QWORD[8+rdi] |
| 1649 | |
| 1650 | lea r12,[r10*2+r14] |
| 1651 | shr r10,63 |
| 1652 | lea r13,[r11*2+rcx] |
| 1653 | shr r11,63 |
| 1654 | or r13,r10 |
| 1655 | mov r10,QWORD[16+rdi] |
| 1656 | mov r14,r11 |
| 1657 | mul rax |
| 1658 | neg r15 |
| 1659 | mov r11,QWORD[24+rdi] |
| 1660 | adc r12,rax |
| 1661 | mov rax,QWORD[((-8))+rbp*1+rsi] |
| 1662 | mov QWORD[rdi],r12 |
| 1663 | adc r13,rdx |
| 1664 | |
| 1665 | lea rbx,[r10*2+r14] |
| 1666 | mov QWORD[8+rdi],r13 |
| 1667 | sbb r15,r15 |
| 1668 | shr r10,63 |
| 1669 | lea r8,[r11*2+rcx] |
| 1670 | shr r11,63 |
| 1671 | or r8,r10 |
| 1672 | mov r10,QWORD[32+rdi] |
| 1673 | mov r14,r11 |
| 1674 | mul rax |
| 1675 | neg r15 |
| 1676 | mov r11,QWORD[40+rdi] |
| 1677 | adc rbx,rax |
| 1678 | mov rax,QWORD[rbp*1+rsi] |
| 1679 | mov QWORD[16+rdi],rbx |
| 1680 | adc r8,rdx |
| 1681 | lea rbp,[16+rbp] |
| 1682 | mov QWORD[24+rdi],r8 |
| 1683 | sbb r15,r15 |
| 1684 | lea rdi,[64+rdi] |
| 1685 | jmp NEAR $L$sqr4x_shift_n_add |
| 1686 | |
| 1687 | ALIGN 32 |
| 1688 | $L$sqr4x_shift_n_add: |
| 1689 | lea r12,[r10*2+r14] |
| 1690 | shr r10,63 |
| 1691 | lea r13,[r11*2+rcx] |
| 1692 | shr r11,63 |
| 1693 | or r13,r10 |
| 1694 | mov r10,QWORD[((-16))+rdi] |
| 1695 | mov r14,r11 |
| 1696 | mul rax |
| 1697 | neg r15 |
| 1698 | mov r11,QWORD[((-8))+rdi] |
| 1699 | adc r12,rax |
| 1700 | mov rax,QWORD[((-8))+rbp*1+rsi] |
| 1701 | mov QWORD[((-32))+rdi],r12 |
| 1702 | adc r13,rdx |
| 1703 | |
| 1704 | lea rbx,[r10*2+r14] |
| 1705 | mov QWORD[((-24))+rdi],r13 |
| 1706 | sbb r15,r15 |
| 1707 | shr r10,63 |
| 1708 | lea r8,[r11*2+rcx] |
| 1709 | shr r11,63 |
| 1710 | or r8,r10 |
| 1711 | mov r10,QWORD[rdi] |
| 1712 | mov r14,r11 |
| 1713 | mul rax |
| 1714 | neg r15 |
| 1715 | mov r11,QWORD[8+rdi] |
| 1716 | adc rbx,rax |
| 1717 | mov rax,QWORD[rbp*1+rsi] |
| 1718 | mov QWORD[((-16))+rdi],rbx |
| 1719 | adc r8,rdx |
| 1720 | |
| 1721 | lea r12,[r10*2+r14] |
| 1722 | mov QWORD[((-8))+rdi],r8 |
| 1723 | sbb r15,r15 |
| 1724 | shr r10,63 |
| 1725 | lea r13,[r11*2+rcx] |
| 1726 | shr r11,63 |
| 1727 | or r13,r10 |
| 1728 | mov r10,QWORD[16+rdi] |
| 1729 | mov r14,r11 |
| 1730 | mul rax |
| 1731 | neg r15 |
| 1732 | mov r11,QWORD[24+rdi] |
| 1733 | adc r12,rax |
| 1734 | mov rax,QWORD[8+rbp*1+rsi] |
| 1735 | mov QWORD[rdi],r12 |
| 1736 | adc r13,rdx |
| 1737 | |
| 1738 | lea rbx,[r10*2+r14] |
| 1739 | mov QWORD[8+rdi],r13 |
| 1740 | sbb r15,r15 |
| 1741 | shr r10,63 |
| 1742 | lea r8,[r11*2+rcx] |
| 1743 | shr r11,63 |
| 1744 | or r8,r10 |
| 1745 | mov r10,QWORD[32+rdi] |
| 1746 | mov r14,r11 |
| 1747 | mul rax |
| 1748 | neg r15 |
| 1749 | mov r11,QWORD[40+rdi] |
| 1750 | adc rbx,rax |
| 1751 | mov rax,QWORD[16+rbp*1+rsi] |
| 1752 | mov QWORD[16+rdi],rbx |
| 1753 | adc r8,rdx |
| 1754 | mov QWORD[24+rdi],r8 |
| 1755 | sbb r15,r15 |
| 1756 | lea rdi,[64+rdi] |
| 1757 | add rbp,32 |
| 1758 | jnz NEAR $L$sqr4x_shift_n_add |
| 1759 | |
| 1760 | lea r12,[r10*2+r14] |
| 1761 | DB 0x67 |
| 1762 | shr r10,63 |
| 1763 | lea r13,[r11*2+rcx] |
| 1764 | shr r11,63 |
| 1765 | or r13,r10 |
| 1766 | mov r10,QWORD[((-16))+rdi] |
| 1767 | mov r14,r11 |
| 1768 | mul rax |
| 1769 | neg r15 |
| 1770 | mov r11,QWORD[((-8))+rdi] |
| 1771 | adc r12,rax |
| 1772 | mov rax,QWORD[((-8))+rsi] |
| 1773 | mov QWORD[((-32))+rdi],r12 |
| 1774 | adc r13,rdx |
| 1775 | |
| 1776 | lea rbx,[r10*2+r14] |
| 1777 | mov QWORD[((-24))+rdi],r13 |
| 1778 | sbb r15,r15 |
| 1779 | shr r10,63 |
| 1780 | lea r8,[r11*2+rcx] |
| 1781 | shr r11,63 |
| 1782 | or r8,r10 |
| 1783 | mul rax |
| 1784 | neg r15 |
| 1785 | adc rbx,rax |
| 1786 | adc r8,rdx |
| 1787 | mov QWORD[((-16))+rdi],rbx |
| 1788 | mov QWORD[((-8))+rdi],r8 |
| 1789 | DB 102,72,15,126,213 |
| 1790 | __bn_sqr8x_reduction: |
| 1791 | xor rax,rax |
| 1792 | lea rcx,[rbp*1+r9] |
| 1793 | lea rdx,[((48+8))+r9*2+rsp] |
| 1794 | mov QWORD[((0+8))+rsp],rcx |
| 1795 | lea rdi,[((48+8))+r9*1+rsp] |
| 1796 | mov QWORD[((8+8))+rsp],rdx |
| 1797 | neg r9 |
| 1798 | jmp NEAR $L$8x_reduction_loop |
| 1799 | |
| 1800 | ALIGN 32 |
| 1801 | $L$8x_reduction_loop: |
| 1802 | lea rdi,[r9*1+rdi] |
| 1803 | DB 0x66 |
| 1804 | mov rbx,QWORD[rdi] |
| 1805 | mov r9,QWORD[8+rdi] |
| 1806 | mov r10,QWORD[16+rdi] |
| 1807 | mov r11,QWORD[24+rdi] |
| 1808 | mov r12,QWORD[32+rdi] |
| 1809 | mov r13,QWORD[40+rdi] |
| 1810 | mov r14,QWORD[48+rdi] |
| 1811 | mov r15,QWORD[56+rdi] |
| 1812 | mov QWORD[rdx],rax |
| 1813 | lea rdi,[64+rdi] |
| 1814 | |
| 1815 | DB 0x67 |
| 1816 | mov r8,rbx |
| 1817 | imul rbx,QWORD[((32+8))+rsp] |
| 1818 | mov rax,QWORD[rbp] |
| 1819 | mov ecx,8 |
| 1820 | jmp NEAR $L$8x_reduce |
| 1821 | |
| 1822 | ALIGN 32 |
| 1823 | $L$8x_reduce: |
| 1824 | mul rbx |
| 1825 | mov rax,QWORD[8+rbp] |
| 1826 | neg r8 |
| 1827 | mov r8,rdx |
| 1828 | adc r8,0 |
| 1829 | |
| 1830 | mul rbx |
| 1831 | add r9,rax |
| 1832 | mov rax,QWORD[16+rbp] |
| 1833 | adc rdx,0 |
| 1834 | add r8,r9 |
| 1835 | mov QWORD[((48-8+8))+rcx*8+rsp],rbx |
| 1836 | mov r9,rdx |
| 1837 | adc r9,0 |
| 1838 | |
| 1839 | mul rbx |
| 1840 | add r10,rax |
| 1841 | mov rax,QWORD[24+rbp] |
| 1842 | adc rdx,0 |
| 1843 | add r9,r10 |
| 1844 | mov rsi,QWORD[((32+8))+rsp] |
| 1845 | mov r10,rdx |
| 1846 | adc r10,0 |
| 1847 | |
| 1848 | mul rbx |
| 1849 | add r11,rax |
| 1850 | mov rax,QWORD[32+rbp] |
| 1851 | adc rdx,0 |
| 1852 | imul rsi,r8 |
| 1853 | add r10,r11 |
| 1854 | mov r11,rdx |
| 1855 | adc r11,0 |
| 1856 | |
| 1857 | mul rbx |
| 1858 | add r12,rax |
| 1859 | mov rax,QWORD[40+rbp] |
| 1860 | adc rdx,0 |
| 1861 | add r11,r12 |
| 1862 | mov r12,rdx |
| 1863 | adc r12,0 |
| 1864 | |
| 1865 | mul rbx |
| 1866 | add r13,rax |
| 1867 | mov rax,QWORD[48+rbp] |
| 1868 | adc rdx,0 |
| 1869 | add r12,r13 |
| 1870 | mov r13,rdx |
| 1871 | adc r13,0 |
| 1872 | |
| 1873 | mul rbx |
| 1874 | add r14,rax |
| 1875 | mov rax,QWORD[56+rbp] |
| 1876 | adc rdx,0 |
| 1877 | add r13,r14 |
| 1878 | mov r14,rdx |
| 1879 | adc r14,0 |
| 1880 | |
| 1881 | mul rbx |
| 1882 | mov rbx,rsi |
| 1883 | add r15,rax |
| 1884 | mov rax,QWORD[rbp] |
| 1885 | adc rdx,0 |
| 1886 | add r14,r15 |
| 1887 | mov r15,rdx |
| 1888 | adc r15,0 |
| 1889 | |
| 1890 | dec ecx |
| 1891 | jnz NEAR $L$8x_reduce |
| 1892 | |
| 1893 | lea rbp,[64+rbp] |
| 1894 | xor rax,rax |
| 1895 | mov rdx,QWORD[((8+8))+rsp] |
| 1896 | cmp rbp,QWORD[((0+8))+rsp] |
| 1897 | jae NEAR $L$8x_no_tail |
| 1898 | |
| 1899 | DB 0x66 |
| 1900 | add r8,QWORD[rdi] |
| 1901 | adc r9,QWORD[8+rdi] |
| 1902 | adc r10,QWORD[16+rdi] |
| 1903 | adc r11,QWORD[24+rdi] |
| 1904 | adc r12,QWORD[32+rdi] |
| 1905 | adc r13,QWORD[40+rdi] |
| 1906 | adc r14,QWORD[48+rdi] |
| 1907 | adc r15,QWORD[56+rdi] |
| 1908 | sbb rsi,rsi |
| 1909 | |
| 1910 | mov rbx,QWORD[((48+56+8))+rsp] |
| 1911 | mov ecx,8 |
| 1912 | mov rax,QWORD[rbp] |
| 1913 | jmp NEAR $L$8x_tail |
| 1914 | |
| 1915 | ALIGN 32 |
| 1916 | $L$8x_tail: |
| 1917 | mul rbx |
| 1918 | add r8,rax |
| 1919 | mov rax,QWORD[8+rbp] |
| 1920 | mov QWORD[rdi],r8 |
| 1921 | mov r8,rdx |
| 1922 | adc r8,0 |
| 1923 | |
| 1924 | mul rbx |
| 1925 | add r9,rax |
| 1926 | mov rax,QWORD[16+rbp] |
| 1927 | adc rdx,0 |
| 1928 | add r8,r9 |
| 1929 | lea rdi,[8+rdi] |
| 1930 | mov r9,rdx |
| 1931 | adc r9,0 |
| 1932 | |
| 1933 | mul rbx |
| 1934 | add r10,rax |
| 1935 | mov rax,QWORD[24+rbp] |
| 1936 | adc rdx,0 |
| 1937 | add r9,r10 |
| 1938 | mov r10,rdx |
| 1939 | adc r10,0 |
| 1940 | |
| 1941 | mul rbx |
| 1942 | add r11,rax |
| 1943 | mov rax,QWORD[32+rbp] |
| 1944 | adc rdx,0 |
| 1945 | add r10,r11 |
| 1946 | mov r11,rdx |
| 1947 | adc r11,0 |
| 1948 | |
| 1949 | mul rbx |
| 1950 | add r12,rax |
| 1951 | mov rax,QWORD[40+rbp] |
| 1952 | adc rdx,0 |
| 1953 | add r11,r12 |
| 1954 | mov r12,rdx |
| 1955 | adc r12,0 |
| 1956 | |
| 1957 | mul rbx |
| 1958 | add r13,rax |
| 1959 | mov rax,QWORD[48+rbp] |
| 1960 | adc rdx,0 |
| 1961 | add r12,r13 |
| 1962 | mov r13,rdx |
| 1963 | adc r13,0 |
| 1964 | |
| 1965 | mul rbx |
| 1966 | add r14,rax |
| 1967 | mov rax,QWORD[56+rbp] |
| 1968 | adc rdx,0 |
| 1969 | add r13,r14 |
| 1970 | mov r14,rdx |
| 1971 | adc r14,0 |
| 1972 | |
| 1973 | mul rbx |
| 1974 | mov rbx,QWORD[((48-16+8))+rcx*8+rsp] |
| 1975 | add r15,rax |
| 1976 | adc rdx,0 |
| 1977 | add r14,r15 |
| 1978 | mov rax,QWORD[rbp] |
| 1979 | mov r15,rdx |
| 1980 | adc r15,0 |
| 1981 | |
| 1982 | dec ecx |
| 1983 | jnz NEAR $L$8x_tail |
| 1984 | |
| 1985 | lea rbp,[64+rbp] |
| 1986 | mov rdx,QWORD[((8+8))+rsp] |
| 1987 | cmp rbp,QWORD[((0+8))+rsp] |
| 1988 | jae NEAR $L$8x_tail_done |
| 1989 | |
| 1990 | mov rbx,QWORD[((48+56+8))+rsp] |
| 1991 | neg rsi |
| 1992 | mov rax,QWORD[rbp] |
| 1993 | adc r8,QWORD[rdi] |
| 1994 | adc r9,QWORD[8+rdi] |
| 1995 | adc r10,QWORD[16+rdi] |
| 1996 | adc r11,QWORD[24+rdi] |
| 1997 | adc r12,QWORD[32+rdi] |
| 1998 | adc r13,QWORD[40+rdi] |
| 1999 | adc r14,QWORD[48+rdi] |
| 2000 | adc r15,QWORD[56+rdi] |
| 2001 | sbb rsi,rsi |
| 2002 | |
| 2003 | mov ecx,8 |
| 2004 | jmp NEAR $L$8x_tail |
| 2005 | |
| 2006 | ALIGN 32 |
| 2007 | $L$8x_tail_done: |
| 2008 | xor rax,rax |
| 2009 | add r8,QWORD[rdx] |
| 2010 | adc r9,0 |
| 2011 | adc r10,0 |
| 2012 | adc r11,0 |
| 2013 | adc r12,0 |
| 2014 | adc r13,0 |
| 2015 | adc r14,0 |
| 2016 | adc r15,0 |
| 2017 | adc rax,0 |
| 2018 | |
| 2019 | neg rsi |
| 2020 | $L$8x_no_tail: |
| 2021 | adc r8,QWORD[rdi] |
| 2022 | adc r9,QWORD[8+rdi] |
| 2023 | adc r10,QWORD[16+rdi] |
| 2024 | adc r11,QWORD[24+rdi] |
| 2025 | adc r12,QWORD[32+rdi] |
| 2026 | adc r13,QWORD[40+rdi] |
| 2027 | adc r14,QWORD[48+rdi] |
| 2028 | adc r15,QWORD[56+rdi] |
| 2029 | adc rax,0 |
| 2030 | mov rcx,QWORD[((-8))+rbp] |
| 2031 | xor rsi,rsi |
| 2032 | |
| 2033 | DB 102,72,15,126,213 |
| 2034 | |
| 2035 | mov QWORD[rdi],r8 |
| 2036 | mov QWORD[8+rdi],r9 |
| 2037 | DB 102,73,15,126,217 |
| 2038 | mov QWORD[16+rdi],r10 |
| 2039 | mov QWORD[24+rdi],r11 |
| 2040 | mov QWORD[32+rdi],r12 |
| 2041 | mov QWORD[40+rdi],r13 |
| 2042 | mov QWORD[48+rdi],r14 |
| 2043 | mov QWORD[56+rdi],r15 |
| 2044 | lea rdi,[64+rdi] |
| 2045 | |
| 2046 | cmp rdi,rdx |
| 2047 | jb NEAR $L$8x_reduction_loop |
| 2048 | ret |
| 2049 | |
| 2050 | |
| 2051 | |
| 2052 | ALIGN 32 |
| 2053 | __bn_post4x_internal: |
| 2054 | |
| 2055 | mov r12,QWORD[rbp] |
| 2056 | lea rbx,[r9*1+rdi] |
| 2057 | mov rcx,r9 |
| 2058 | DB 102,72,15,126,207 |
| 2059 | neg rax |
| 2060 | DB 102,72,15,126,206 |
| 2061 | sar rcx,3+2 |
| 2062 | dec r12 |
| 2063 | xor r10,r10 |
| 2064 | mov r13,QWORD[8+rbp] |
| 2065 | mov r14,QWORD[16+rbp] |
| 2066 | mov r15,QWORD[24+rbp] |
| 2067 | jmp NEAR $L$sqr4x_sub_entry |
| 2068 | |
| 2069 | ALIGN 16 |
| 2070 | $L$sqr4x_sub: |
| 2071 | mov r12,QWORD[rbp] |
| 2072 | mov r13,QWORD[8+rbp] |
| 2073 | mov r14,QWORD[16+rbp] |
| 2074 | mov r15,QWORD[24+rbp] |
| 2075 | $L$sqr4x_sub_entry: |
| 2076 | lea rbp,[32+rbp] |
| 2077 | not r12 |
| 2078 | not r13 |
| 2079 | not r14 |
| 2080 | not r15 |
| 2081 | and r12,rax |
| 2082 | and r13,rax |
| 2083 | and r14,rax |
| 2084 | and r15,rax |
| 2085 | |
| 2086 | neg r10 |
| 2087 | adc r12,QWORD[rbx] |
| 2088 | adc r13,QWORD[8+rbx] |
| 2089 | adc r14,QWORD[16+rbx] |
| 2090 | adc r15,QWORD[24+rbx] |
| 2091 | mov QWORD[rdi],r12 |
| 2092 | lea rbx,[32+rbx] |
| 2093 | mov QWORD[8+rdi],r13 |
| 2094 | sbb r10,r10 |
| 2095 | mov QWORD[16+rdi],r14 |
| 2096 | mov QWORD[24+rdi],r15 |
| 2097 | lea rdi,[32+rdi] |
| 2098 | |
| 2099 | inc rcx |
| 2100 | jnz NEAR $L$sqr4x_sub |
| 2101 | |
| 2102 | mov r10,r9 |
| 2103 | neg r9 |
| 2104 | ret |
| 2105 | |
| 2106 | |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 2107 | global bn_mulx4x_mont_gather5 |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 2108 | |
| 2109 | ALIGN 32 |
| 2110 | bn_mulx4x_mont_gather5: |
| 2111 | mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 2112 | mov QWORD[16+rsp],rsi |
| 2113 | mov rax,rsp |
| 2114 | $L$SEH_begin_bn_mulx4x_mont_gather5: |
| 2115 | mov rdi,rcx |
| 2116 | mov rsi,rdx |
| 2117 | mov rdx,r8 |
| 2118 | mov rcx,r9 |
| 2119 | mov r8,QWORD[40+rsp] |
| 2120 | mov r9,QWORD[48+rsp] |
| 2121 | |
| 2122 | |
| 2123 | |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 2124 | _CET_ENDBR |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 2125 | mov rax,rsp |
| 2126 | |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 2127 | push rbx |
| 2128 | |
| 2129 | push rbp |
| 2130 | |
| 2131 | push r12 |
| 2132 | |
| 2133 | push r13 |
| 2134 | |
| 2135 | push r14 |
| 2136 | |
| 2137 | push r15 |
| 2138 | |
| 2139 | $L$mulx4x_prologue: |
| 2140 | |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 2141 | |
| 2142 | |
| 2143 | |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 2144 | shl r9d,3 |
| 2145 | lea r10,[r9*2+r9] |
| 2146 | neg r9 |
| 2147 | mov r8,QWORD[r8] |
| 2148 | |
| 2149 | |
| 2150 | |
| 2151 | |
| 2152 | |
| 2153 | |
| 2154 | |
| 2155 | |
| 2156 | |
| 2157 | |
| 2158 | lea r11,[((-320))+r9*2+rsp] |
| 2159 | mov rbp,rsp |
| 2160 | sub r11,rdi |
| 2161 | and r11,4095 |
| 2162 | cmp r10,r11 |
| 2163 | jb NEAR $L$mulx4xsp_alt |
| 2164 | sub rbp,r11 |
| 2165 | lea rbp,[((-320))+r9*2+rbp] |
| 2166 | jmp NEAR $L$mulx4xsp_done |
| 2167 | |
| 2168 | $L$mulx4xsp_alt: |
| 2169 | lea r10,[((4096-320))+r9*2] |
| 2170 | lea rbp,[((-320))+r9*2+rbp] |
| 2171 | sub r11,r10 |
| 2172 | mov r10,0 |
| 2173 | cmovc r11,r10 |
| 2174 | sub rbp,r11 |
| 2175 | $L$mulx4xsp_done: |
| 2176 | and rbp,-64 |
| 2177 | mov r11,rsp |
| 2178 | sub r11,rbp |
| 2179 | and r11,-4096 |
| 2180 | lea rsp,[rbp*1+r11] |
| 2181 | mov r10,QWORD[rsp] |
| 2182 | cmp rsp,rbp |
| 2183 | ja NEAR $L$mulx4x_page_walk |
| 2184 | jmp NEAR $L$mulx4x_page_walk_done |
| 2185 | |
| 2186 | $L$mulx4x_page_walk: |
| 2187 | lea rsp,[((-4096))+rsp] |
| 2188 | mov r10,QWORD[rsp] |
| 2189 | cmp rsp,rbp |
| 2190 | ja NEAR $L$mulx4x_page_walk |
| 2191 | $L$mulx4x_page_walk_done: |
| 2192 | |
| 2193 | |
| 2194 | |
| 2195 | |
| 2196 | |
| 2197 | |
| 2198 | |
| 2199 | |
| 2200 | |
| 2201 | |
| 2202 | |
| 2203 | |
| 2204 | |
| 2205 | mov QWORD[32+rsp],r8 |
| 2206 | mov QWORD[40+rsp],rax |
| 2207 | |
| 2208 | $L$mulx4x_body: |
| 2209 | call mulx4x_internal |
| 2210 | |
| 2211 | mov rsi,QWORD[40+rsp] |
| 2212 | |
| 2213 | mov rax,1 |
| 2214 | |
| 2215 | mov r15,QWORD[((-48))+rsi] |
| 2216 | |
| 2217 | mov r14,QWORD[((-40))+rsi] |
| 2218 | |
| 2219 | mov r13,QWORD[((-32))+rsi] |
| 2220 | |
| 2221 | mov r12,QWORD[((-24))+rsi] |
| 2222 | |
| 2223 | mov rbp,QWORD[((-16))+rsi] |
| 2224 | |
| 2225 | mov rbx,QWORD[((-8))+rsi] |
| 2226 | |
| 2227 | lea rsp,[rsi] |
| 2228 | |
| 2229 | $L$mulx4x_epilogue: |
| 2230 | mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 2231 | mov rsi,QWORD[16+rsp] |
| 2232 | ret |
| 2233 | |
| 2234 | $L$SEH_end_bn_mulx4x_mont_gather5: |
| 2235 | |
| 2236 | |
| 2237 | ALIGN 32 |
| 2238 | mulx4x_internal: |
| 2239 | |
| 2240 | mov QWORD[8+rsp],r9 |
| 2241 | mov r10,r9 |
| 2242 | neg r9 |
| 2243 | shl r9,5 |
| 2244 | neg r10 |
| 2245 | lea r13,[128+r9*1+rdx] |
| 2246 | shr r9,5+5 |
| 2247 | movd xmm5,DWORD[56+rax] |
| 2248 | sub r9,1 |
| 2249 | lea rax,[$L$inc] |
| 2250 | mov QWORD[((16+8))+rsp],r13 |
| 2251 | mov QWORD[((24+8))+rsp],r9 |
| 2252 | mov QWORD[((56+8))+rsp],rdi |
| 2253 | movdqa xmm0,XMMWORD[rax] |
| 2254 | movdqa xmm1,XMMWORD[16+rax] |
| 2255 | lea r10,[((88-112))+r10*1+rsp] |
| 2256 | lea rdi,[128+rdx] |
| 2257 | |
| 2258 | pshufd xmm5,xmm5,0 |
| 2259 | movdqa xmm4,xmm1 |
| 2260 | DB 0x67 |
| 2261 | movdqa xmm2,xmm1 |
| 2262 | DB 0x67 |
| 2263 | paddd xmm1,xmm0 |
| 2264 | pcmpeqd xmm0,xmm5 |
| 2265 | movdqa xmm3,xmm4 |
| 2266 | paddd xmm2,xmm1 |
| 2267 | pcmpeqd xmm1,xmm5 |
| 2268 | movdqa XMMWORD[112+r10],xmm0 |
| 2269 | movdqa xmm0,xmm4 |
| 2270 | |
| 2271 | paddd xmm3,xmm2 |
| 2272 | pcmpeqd xmm2,xmm5 |
| 2273 | movdqa XMMWORD[128+r10],xmm1 |
| 2274 | movdqa xmm1,xmm4 |
| 2275 | |
| 2276 | paddd xmm0,xmm3 |
| 2277 | pcmpeqd xmm3,xmm5 |
| 2278 | movdqa XMMWORD[144+r10],xmm2 |
| 2279 | movdqa xmm2,xmm4 |
| 2280 | |
| 2281 | paddd xmm1,xmm0 |
| 2282 | pcmpeqd xmm0,xmm5 |
| 2283 | movdqa XMMWORD[160+r10],xmm3 |
| 2284 | movdqa xmm3,xmm4 |
| 2285 | paddd xmm2,xmm1 |
| 2286 | pcmpeqd xmm1,xmm5 |
| 2287 | movdqa XMMWORD[176+r10],xmm0 |
| 2288 | movdqa xmm0,xmm4 |
| 2289 | |
| 2290 | paddd xmm3,xmm2 |
| 2291 | pcmpeqd xmm2,xmm5 |
| 2292 | movdqa XMMWORD[192+r10],xmm1 |
| 2293 | movdqa xmm1,xmm4 |
| 2294 | |
| 2295 | paddd xmm0,xmm3 |
| 2296 | pcmpeqd xmm3,xmm5 |
| 2297 | movdqa XMMWORD[208+r10],xmm2 |
| 2298 | movdqa xmm2,xmm4 |
| 2299 | |
| 2300 | paddd xmm1,xmm0 |
| 2301 | pcmpeqd xmm0,xmm5 |
| 2302 | movdqa XMMWORD[224+r10],xmm3 |
| 2303 | movdqa xmm3,xmm4 |
| 2304 | paddd xmm2,xmm1 |
| 2305 | pcmpeqd xmm1,xmm5 |
| 2306 | movdqa XMMWORD[240+r10],xmm0 |
| 2307 | movdqa xmm0,xmm4 |
| 2308 | |
| 2309 | paddd xmm3,xmm2 |
| 2310 | pcmpeqd xmm2,xmm5 |
| 2311 | movdqa XMMWORD[256+r10],xmm1 |
| 2312 | movdqa xmm1,xmm4 |
| 2313 | |
| 2314 | paddd xmm0,xmm3 |
| 2315 | pcmpeqd xmm3,xmm5 |
| 2316 | movdqa XMMWORD[272+r10],xmm2 |
| 2317 | movdqa xmm2,xmm4 |
| 2318 | |
| 2319 | paddd xmm1,xmm0 |
| 2320 | pcmpeqd xmm0,xmm5 |
| 2321 | movdqa XMMWORD[288+r10],xmm3 |
| 2322 | movdqa xmm3,xmm4 |
| 2323 | DB 0x67 |
| 2324 | paddd xmm2,xmm1 |
| 2325 | pcmpeqd xmm1,xmm5 |
| 2326 | movdqa XMMWORD[304+r10],xmm0 |
| 2327 | |
| 2328 | paddd xmm3,xmm2 |
| 2329 | pcmpeqd xmm2,xmm5 |
| 2330 | movdqa XMMWORD[320+r10],xmm1 |
| 2331 | |
| 2332 | pcmpeqd xmm3,xmm5 |
| 2333 | movdqa XMMWORD[336+r10],xmm2 |
| 2334 | |
| 2335 | pand xmm0,XMMWORD[64+rdi] |
| 2336 | pand xmm1,XMMWORD[80+rdi] |
| 2337 | pand xmm2,XMMWORD[96+rdi] |
| 2338 | movdqa XMMWORD[352+r10],xmm3 |
| 2339 | pand xmm3,XMMWORD[112+rdi] |
| 2340 | por xmm0,xmm2 |
| 2341 | por xmm1,xmm3 |
| 2342 | movdqa xmm4,XMMWORD[((-128))+rdi] |
| 2343 | movdqa xmm5,XMMWORD[((-112))+rdi] |
| 2344 | movdqa xmm2,XMMWORD[((-96))+rdi] |
| 2345 | pand xmm4,XMMWORD[112+r10] |
| 2346 | movdqa xmm3,XMMWORD[((-80))+rdi] |
| 2347 | pand xmm5,XMMWORD[128+r10] |
| 2348 | por xmm0,xmm4 |
| 2349 | pand xmm2,XMMWORD[144+r10] |
| 2350 | por xmm1,xmm5 |
| 2351 | pand xmm3,XMMWORD[160+r10] |
| 2352 | por xmm0,xmm2 |
| 2353 | por xmm1,xmm3 |
| 2354 | movdqa xmm4,XMMWORD[((-64))+rdi] |
| 2355 | movdqa xmm5,XMMWORD[((-48))+rdi] |
| 2356 | movdqa xmm2,XMMWORD[((-32))+rdi] |
| 2357 | pand xmm4,XMMWORD[176+r10] |
| 2358 | movdqa xmm3,XMMWORD[((-16))+rdi] |
| 2359 | pand xmm5,XMMWORD[192+r10] |
| 2360 | por xmm0,xmm4 |
| 2361 | pand xmm2,XMMWORD[208+r10] |
| 2362 | por xmm1,xmm5 |
| 2363 | pand xmm3,XMMWORD[224+r10] |
| 2364 | por xmm0,xmm2 |
| 2365 | por xmm1,xmm3 |
| 2366 | movdqa xmm4,XMMWORD[rdi] |
| 2367 | movdqa xmm5,XMMWORD[16+rdi] |
| 2368 | movdqa xmm2,XMMWORD[32+rdi] |
| 2369 | pand xmm4,XMMWORD[240+r10] |
| 2370 | movdqa xmm3,XMMWORD[48+rdi] |
| 2371 | pand xmm5,XMMWORD[256+r10] |
| 2372 | por xmm0,xmm4 |
| 2373 | pand xmm2,XMMWORD[272+r10] |
| 2374 | por xmm1,xmm5 |
| 2375 | pand xmm3,XMMWORD[288+r10] |
| 2376 | por xmm0,xmm2 |
| 2377 | por xmm1,xmm3 |
| 2378 | pxor xmm0,xmm1 |
| 2379 | |
| 2380 | pshufd xmm1,xmm0,0x4e |
| 2381 | por xmm0,xmm1 |
| 2382 | lea rdi,[256+rdi] |
| 2383 | DB 102,72,15,126,194 |
| 2384 | lea rbx,[((64+32+8))+rsp] |
| 2385 | |
| 2386 | mov r9,rdx |
| 2387 | mulx rax,r8,QWORD[rsi] |
| 2388 | mulx r12,r11,QWORD[8+rsi] |
| 2389 | add r11,rax |
| 2390 | mulx r13,rax,QWORD[16+rsi] |
| 2391 | adc r12,rax |
| 2392 | adc r13,0 |
| 2393 | mulx r14,rax,QWORD[24+rsi] |
| 2394 | |
| 2395 | mov r15,r8 |
| 2396 | imul r8,QWORD[((32+8))+rsp] |
| 2397 | xor rbp,rbp |
| 2398 | mov rdx,r8 |
| 2399 | |
| 2400 | mov QWORD[((8+8))+rsp],rdi |
| 2401 | |
| 2402 | lea rsi,[32+rsi] |
| 2403 | adcx r13,rax |
| 2404 | adcx r14,rbp |
| 2405 | |
| 2406 | mulx r10,rax,QWORD[rcx] |
| 2407 | adcx r15,rax |
| 2408 | adox r10,r11 |
| 2409 | mulx r11,rax,QWORD[8+rcx] |
| 2410 | adcx r10,rax |
| 2411 | adox r11,r12 |
| 2412 | mulx r12,rax,QWORD[16+rcx] |
| 2413 | mov rdi,QWORD[((24+8))+rsp] |
| 2414 | mov QWORD[((-32))+rbx],r10 |
| 2415 | adcx r11,rax |
| 2416 | adox r12,r13 |
| 2417 | mulx r15,rax,QWORD[24+rcx] |
| 2418 | mov rdx,r9 |
| 2419 | mov QWORD[((-24))+rbx],r11 |
| 2420 | adcx r12,rax |
| 2421 | adox r15,rbp |
| 2422 | lea rcx,[32+rcx] |
| 2423 | mov QWORD[((-16))+rbx],r12 |
| 2424 | jmp NEAR $L$mulx4x_1st |
| 2425 | |
| 2426 | ALIGN 32 |
| 2427 | $L$mulx4x_1st: |
| 2428 | adcx r15,rbp |
| 2429 | mulx rax,r10,QWORD[rsi] |
| 2430 | adcx r10,r14 |
| 2431 | mulx r14,r11,QWORD[8+rsi] |
| 2432 | adcx r11,rax |
| 2433 | mulx rax,r12,QWORD[16+rsi] |
| 2434 | adcx r12,r14 |
| 2435 | mulx r14,r13,QWORD[24+rsi] |
| 2436 | DB 0x67,0x67 |
| 2437 | mov rdx,r8 |
| 2438 | adcx r13,rax |
| 2439 | adcx r14,rbp |
| 2440 | lea rsi,[32+rsi] |
| 2441 | lea rbx,[32+rbx] |
| 2442 | |
| 2443 | adox r10,r15 |
| 2444 | mulx r15,rax,QWORD[rcx] |
| 2445 | adcx r10,rax |
| 2446 | adox r11,r15 |
| 2447 | mulx r15,rax,QWORD[8+rcx] |
| 2448 | adcx r11,rax |
| 2449 | adox r12,r15 |
| 2450 | mulx r15,rax,QWORD[16+rcx] |
| 2451 | mov QWORD[((-40))+rbx],r10 |
| 2452 | adcx r12,rax |
| 2453 | mov QWORD[((-32))+rbx],r11 |
| 2454 | adox r13,r15 |
| 2455 | mulx r15,rax,QWORD[24+rcx] |
| 2456 | mov rdx,r9 |
| 2457 | mov QWORD[((-24))+rbx],r12 |
| 2458 | adcx r13,rax |
| 2459 | adox r15,rbp |
| 2460 | lea rcx,[32+rcx] |
| 2461 | mov QWORD[((-16))+rbx],r13 |
| 2462 | |
| 2463 | dec rdi |
| 2464 | jnz NEAR $L$mulx4x_1st |
| 2465 | |
| 2466 | mov rax,QWORD[8+rsp] |
| 2467 | adc r15,rbp |
| 2468 | lea rsi,[rax*1+rsi] |
| 2469 | add r14,r15 |
| 2470 | mov rdi,QWORD[((8+8))+rsp] |
| 2471 | adc rbp,rbp |
| 2472 | mov QWORD[((-8))+rbx],r14 |
| 2473 | jmp NEAR $L$mulx4x_outer |
| 2474 | |
| 2475 | ALIGN 32 |
| 2476 | $L$mulx4x_outer: |
| 2477 | lea r10,[((16-256))+rbx] |
| 2478 | pxor xmm4,xmm4 |
| 2479 | DB 0x67,0x67 |
| 2480 | pxor xmm5,xmm5 |
| 2481 | movdqa xmm0,XMMWORD[((-128))+rdi] |
| 2482 | movdqa xmm1,XMMWORD[((-112))+rdi] |
| 2483 | movdqa xmm2,XMMWORD[((-96))+rdi] |
| 2484 | pand xmm0,XMMWORD[256+r10] |
| 2485 | movdqa xmm3,XMMWORD[((-80))+rdi] |
| 2486 | pand xmm1,XMMWORD[272+r10] |
| 2487 | por xmm4,xmm0 |
| 2488 | pand xmm2,XMMWORD[288+r10] |
| 2489 | por xmm5,xmm1 |
| 2490 | pand xmm3,XMMWORD[304+r10] |
| 2491 | por xmm4,xmm2 |
| 2492 | por xmm5,xmm3 |
| 2493 | movdqa xmm0,XMMWORD[((-64))+rdi] |
| 2494 | movdqa xmm1,XMMWORD[((-48))+rdi] |
| 2495 | movdqa xmm2,XMMWORD[((-32))+rdi] |
| 2496 | pand xmm0,XMMWORD[320+r10] |
| 2497 | movdqa xmm3,XMMWORD[((-16))+rdi] |
| 2498 | pand xmm1,XMMWORD[336+r10] |
| 2499 | por xmm4,xmm0 |
| 2500 | pand xmm2,XMMWORD[352+r10] |
| 2501 | por xmm5,xmm1 |
| 2502 | pand xmm3,XMMWORD[368+r10] |
| 2503 | por xmm4,xmm2 |
| 2504 | por xmm5,xmm3 |
| 2505 | movdqa xmm0,XMMWORD[rdi] |
| 2506 | movdqa xmm1,XMMWORD[16+rdi] |
| 2507 | movdqa xmm2,XMMWORD[32+rdi] |
| 2508 | pand xmm0,XMMWORD[384+r10] |
| 2509 | movdqa xmm3,XMMWORD[48+rdi] |
| 2510 | pand xmm1,XMMWORD[400+r10] |
| 2511 | por xmm4,xmm0 |
| 2512 | pand xmm2,XMMWORD[416+r10] |
| 2513 | por xmm5,xmm1 |
| 2514 | pand xmm3,XMMWORD[432+r10] |
| 2515 | por xmm4,xmm2 |
| 2516 | por xmm5,xmm3 |
| 2517 | movdqa xmm0,XMMWORD[64+rdi] |
| 2518 | movdqa xmm1,XMMWORD[80+rdi] |
| 2519 | movdqa xmm2,XMMWORD[96+rdi] |
| 2520 | pand xmm0,XMMWORD[448+r10] |
| 2521 | movdqa xmm3,XMMWORD[112+rdi] |
| 2522 | pand xmm1,XMMWORD[464+r10] |
| 2523 | por xmm4,xmm0 |
| 2524 | pand xmm2,XMMWORD[480+r10] |
| 2525 | por xmm5,xmm1 |
| 2526 | pand xmm3,XMMWORD[496+r10] |
| 2527 | por xmm4,xmm2 |
| 2528 | por xmm5,xmm3 |
| 2529 | por xmm4,xmm5 |
| 2530 | |
| 2531 | pshufd xmm0,xmm4,0x4e |
| 2532 | por xmm0,xmm4 |
| 2533 | lea rdi,[256+rdi] |
| 2534 | DB 102,72,15,126,194 |
| 2535 | |
| 2536 | mov QWORD[rbx],rbp |
| 2537 | lea rbx,[32+rax*1+rbx] |
| 2538 | mulx r11,r8,QWORD[rsi] |
| 2539 | xor rbp,rbp |
| 2540 | mov r9,rdx |
| 2541 | mulx r12,r14,QWORD[8+rsi] |
| 2542 | adox r8,QWORD[((-32))+rbx] |
| 2543 | adcx r11,r14 |
| 2544 | mulx r13,r15,QWORD[16+rsi] |
| 2545 | adox r11,QWORD[((-24))+rbx] |
| 2546 | adcx r12,r15 |
| 2547 | mulx r14,rdx,QWORD[24+rsi] |
| 2548 | adox r12,QWORD[((-16))+rbx] |
| 2549 | adcx r13,rdx |
| 2550 | lea rcx,[rax*1+rcx] |
| 2551 | lea rsi,[32+rsi] |
| 2552 | adox r13,QWORD[((-8))+rbx] |
| 2553 | adcx r14,rbp |
| 2554 | adox r14,rbp |
| 2555 | |
| 2556 | mov r15,r8 |
| 2557 | imul r8,QWORD[((32+8))+rsp] |
| 2558 | |
| 2559 | mov rdx,r8 |
| 2560 | xor rbp,rbp |
| 2561 | mov QWORD[((8+8))+rsp],rdi |
| 2562 | |
| 2563 | mulx r10,rax,QWORD[rcx] |
| 2564 | adcx r15,rax |
| 2565 | adox r10,r11 |
| 2566 | mulx r11,rax,QWORD[8+rcx] |
| 2567 | adcx r10,rax |
| 2568 | adox r11,r12 |
| 2569 | mulx r12,rax,QWORD[16+rcx] |
| 2570 | adcx r11,rax |
| 2571 | adox r12,r13 |
| 2572 | mulx r15,rax,QWORD[24+rcx] |
| 2573 | mov rdx,r9 |
| 2574 | mov rdi,QWORD[((24+8))+rsp] |
| 2575 | mov QWORD[((-32))+rbx],r10 |
| 2576 | adcx r12,rax |
| 2577 | mov QWORD[((-24))+rbx],r11 |
| 2578 | adox r15,rbp |
| 2579 | mov QWORD[((-16))+rbx],r12 |
| 2580 | lea rcx,[32+rcx] |
| 2581 | jmp NEAR $L$mulx4x_inner |
| 2582 | |
| 2583 | ALIGN 32 |
| 2584 | $L$mulx4x_inner: |
| 2585 | mulx rax,r10,QWORD[rsi] |
| 2586 | adcx r15,rbp |
| 2587 | adox r10,r14 |
| 2588 | mulx r14,r11,QWORD[8+rsi] |
| 2589 | adcx r10,QWORD[rbx] |
| 2590 | adox r11,rax |
| 2591 | mulx rax,r12,QWORD[16+rsi] |
| 2592 | adcx r11,QWORD[8+rbx] |
| 2593 | adox r12,r14 |
| 2594 | mulx r14,r13,QWORD[24+rsi] |
| 2595 | mov rdx,r8 |
| 2596 | adcx r12,QWORD[16+rbx] |
| 2597 | adox r13,rax |
| 2598 | adcx r13,QWORD[24+rbx] |
| 2599 | adox r14,rbp |
| 2600 | lea rsi,[32+rsi] |
| 2601 | lea rbx,[32+rbx] |
| 2602 | adcx r14,rbp |
| 2603 | |
| 2604 | adox r10,r15 |
| 2605 | mulx r15,rax,QWORD[rcx] |
| 2606 | adcx r10,rax |
| 2607 | adox r11,r15 |
| 2608 | mulx r15,rax,QWORD[8+rcx] |
| 2609 | adcx r11,rax |
| 2610 | adox r12,r15 |
| 2611 | mulx r15,rax,QWORD[16+rcx] |
| 2612 | mov QWORD[((-40))+rbx],r10 |
| 2613 | adcx r12,rax |
| 2614 | adox r13,r15 |
| 2615 | mov QWORD[((-32))+rbx],r11 |
| 2616 | mulx r15,rax,QWORD[24+rcx] |
| 2617 | mov rdx,r9 |
| 2618 | lea rcx,[32+rcx] |
| 2619 | mov QWORD[((-24))+rbx],r12 |
| 2620 | adcx r13,rax |
| 2621 | adox r15,rbp |
| 2622 | mov QWORD[((-16))+rbx],r13 |
| 2623 | |
| 2624 | dec rdi |
| 2625 | jnz NEAR $L$mulx4x_inner |
| 2626 | |
| 2627 | mov rax,QWORD[((0+8))+rsp] |
| 2628 | adc r15,rbp |
| 2629 | sub rdi,QWORD[rbx] |
| 2630 | mov rdi,QWORD[((8+8))+rsp] |
| 2631 | mov r10,QWORD[((16+8))+rsp] |
| 2632 | adc r14,r15 |
| 2633 | lea rsi,[rax*1+rsi] |
| 2634 | adc rbp,rbp |
| 2635 | mov QWORD[((-8))+rbx],r14 |
| 2636 | |
| 2637 | cmp rdi,r10 |
| 2638 | jb NEAR $L$mulx4x_outer |
| 2639 | |
| 2640 | mov r10,QWORD[((-8))+rcx] |
| 2641 | mov r8,rbp |
| 2642 | mov r12,QWORD[rax*1+rcx] |
| 2643 | lea rbp,[rax*1+rcx] |
| 2644 | mov rcx,rax |
| 2645 | lea rdi,[rax*1+rbx] |
| 2646 | xor eax,eax |
| 2647 | xor r15,r15 |
| 2648 | sub r10,r14 |
| 2649 | adc r15,r15 |
| 2650 | or r8,r15 |
| 2651 | sar rcx,3+2 |
| 2652 | sub rax,r8 |
| 2653 | mov rdx,QWORD[((56+8))+rsp] |
| 2654 | dec r12 |
| 2655 | mov r13,QWORD[8+rbp] |
| 2656 | xor r8,r8 |
| 2657 | mov r14,QWORD[16+rbp] |
| 2658 | mov r15,QWORD[24+rbp] |
| 2659 | jmp NEAR $L$sqrx4x_sub_entry |
| 2660 | |
| 2661 | |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 2662 | global bn_powerx5 |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 2663 | |
| 2664 | ALIGN 32 |
| 2665 | bn_powerx5: |
| 2666 | mov QWORD[8+rsp],rdi ;WIN64 prologue |
| 2667 | mov QWORD[16+rsp],rsi |
| 2668 | mov rax,rsp |
| 2669 | $L$SEH_begin_bn_powerx5: |
| 2670 | mov rdi,rcx |
| 2671 | mov rsi,rdx |
| 2672 | mov rdx,r8 |
| 2673 | mov rcx,r9 |
| 2674 | mov r8,QWORD[40+rsp] |
| 2675 | mov r9,QWORD[48+rsp] |
| 2676 | |
| 2677 | |
| 2678 | |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 2679 | _CET_ENDBR |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 2680 | mov rax,rsp |
| 2681 | |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 2682 | push rbx |
| 2683 | |
| 2684 | push rbp |
| 2685 | |
| 2686 | push r12 |
| 2687 | |
| 2688 | push r13 |
| 2689 | |
| 2690 | push r14 |
| 2691 | |
| 2692 | push r15 |
| 2693 | |
| 2694 | $L$powerx5_prologue: |
| 2695 | |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 2696 | |
| 2697 | |
| 2698 | |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 2699 | shl r9d,3 |
| 2700 | lea r10,[r9*2+r9] |
| 2701 | neg r9 |
| 2702 | mov r8,QWORD[r8] |
| 2703 | |
| 2704 | |
| 2705 | |
| 2706 | |
| 2707 | |
| 2708 | |
| 2709 | |
| 2710 | |
| 2711 | lea r11,[((-320))+r9*2+rsp] |
| 2712 | mov rbp,rsp |
| 2713 | sub r11,rdi |
| 2714 | and r11,4095 |
| 2715 | cmp r10,r11 |
| 2716 | jb NEAR $L$pwrx_sp_alt |
| 2717 | sub rbp,r11 |
| 2718 | lea rbp,[((-320))+r9*2+rbp] |
| 2719 | jmp NEAR $L$pwrx_sp_done |
| 2720 | |
| 2721 | ALIGN 32 |
| 2722 | $L$pwrx_sp_alt: |
| 2723 | lea r10,[((4096-320))+r9*2] |
| 2724 | lea rbp,[((-320))+r9*2+rbp] |
| 2725 | sub r11,r10 |
| 2726 | mov r10,0 |
| 2727 | cmovc r11,r10 |
| 2728 | sub rbp,r11 |
| 2729 | $L$pwrx_sp_done: |
| 2730 | and rbp,-64 |
| 2731 | mov r11,rsp |
| 2732 | sub r11,rbp |
| 2733 | and r11,-4096 |
| 2734 | lea rsp,[rbp*1+r11] |
| 2735 | mov r10,QWORD[rsp] |
| 2736 | cmp rsp,rbp |
| 2737 | ja NEAR $L$pwrx_page_walk |
| 2738 | jmp NEAR $L$pwrx_page_walk_done |
| 2739 | |
| 2740 | $L$pwrx_page_walk: |
| 2741 | lea rsp,[((-4096))+rsp] |
| 2742 | mov r10,QWORD[rsp] |
| 2743 | cmp rsp,rbp |
| 2744 | ja NEAR $L$pwrx_page_walk |
| 2745 | $L$pwrx_page_walk_done: |
| 2746 | |
| 2747 | mov r10,r9 |
| 2748 | neg r9 |
| 2749 | |
| 2750 | |
| 2751 | |
| 2752 | |
| 2753 | |
| 2754 | |
| 2755 | |
| 2756 | |
| 2757 | |
| 2758 | |
| 2759 | |
| 2760 | |
| 2761 | pxor xmm0,xmm0 |
| 2762 | DB 102,72,15,110,207 |
| 2763 | DB 102,72,15,110,209 |
| 2764 | DB 102,73,15,110,218 |
| 2765 | DB 102,72,15,110,226 |
| 2766 | mov QWORD[32+rsp],r8 |
| 2767 | mov QWORD[40+rsp],rax |
| 2768 | |
| 2769 | $L$powerx5_body: |
| 2770 | |
| 2771 | call __bn_sqrx8x_internal |
| 2772 | call __bn_postx4x_internal |
| 2773 | call __bn_sqrx8x_internal |
| 2774 | call __bn_postx4x_internal |
| 2775 | call __bn_sqrx8x_internal |
| 2776 | call __bn_postx4x_internal |
| 2777 | call __bn_sqrx8x_internal |
| 2778 | call __bn_postx4x_internal |
| 2779 | call __bn_sqrx8x_internal |
| 2780 | call __bn_postx4x_internal |
| 2781 | |
| 2782 | mov r9,r10 |
| 2783 | mov rdi,rsi |
| 2784 | DB 102,72,15,126,209 |
| 2785 | DB 102,72,15,126,226 |
| 2786 | mov rax,QWORD[40+rsp] |
| 2787 | |
| 2788 | call mulx4x_internal |
| 2789 | |
| 2790 | mov rsi,QWORD[40+rsp] |
| 2791 | |
| 2792 | mov rax,1 |
| 2793 | |
| 2794 | mov r15,QWORD[((-48))+rsi] |
| 2795 | |
| 2796 | mov r14,QWORD[((-40))+rsi] |
| 2797 | |
| 2798 | mov r13,QWORD[((-32))+rsi] |
| 2799 | |
| 2800 | mov r12,QWORD[((-24))+rsi] |
| 2801 | |
| 2802 | mov rbp,QWORD[((-16))+rsi] |
| 2803 | |
| 2804 | mov rbx,QWORD[((-8))+rsi] |
| 2805 | |
| 2806 | lea rsp,[rsi] |
| 2807 | |
| 2808 | $L$powerx5_epilogue: |
| 2809 | mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| 2810 | mov rsi,QWORD[16+rsp] |
| 2811 | ret |
| 2812 | |
| 2813 | $L$SEH_end_bn_powerx5: |
| 2814 | |
| 2815 | global bn_sqrx8x_internal |
| 2816 | |
| 2817 | |
| 2818 | ALIGN 32 |
| 2819 | bn_sqrx8x_internal: |
| 2820 | __bn_sqrx8x_internal: |
| 2821 | |
| 2822 | _CET_ENDBR |
| 2823 | |
| 2824 | |
| 2825 | |
| 2826 | |
| 2827 | |
| 2828 | |
| 2829 | |
| 2830 | |
| 2831 | |
| 2832 | |
| 2833 | |
| 2834 | |
| 2835 | |
| 2836 | |
| 2837 | |
| 2838 | |
| 2839 | |
| 2840 | |
| 2841 | |
| 2842 | |
| 2843 | |
| 2844 | |
| 2845 | |
| 2846 | |
| 2847 | |
| 2848 | |
| 2849 | |
| 2850 | |
| 2851 | |
| 2852 | |
| 2853 | |
| 2854 | |
| 2855 | |
| 2856 | |
| 2857 | |
| 2858 | |
| 2859 | |
| 2860 | |
| 2861 | |
| 2862 | |
| 2863 | lea rdi,[((48+8))+rsp] |
| 2864 | lea rbp,[r9*1+rsi] |
| 2865 | mov QWORD[((0+8))+rsp],r9 |
| 2866 | mov QWORD[((8+8))+rsp],rbp |
| 2867 | jmp NEAR $L$sqr8x_zero_start |
| 2868 | |
| 2869 | ALIGN 32 |
| 2870 | DB 0x66,0x66,0x66,0x2e,0x0f,0x1f,0x84,0x00,0x00,0x00,0x00,0x00 |
| 2871 | $L$sqrx8x_zero: |
| 2872 | DB 0x3e |
| 2873 | movdqa XMMWORD[rdi],xmm0 |
| 2874 | movdqa XMMWORD[16+rdi],xmm0 |
| 2875 | movdqa XMMWORD[32+rdi],xmm0 |
| 2876 | movdqa XMMWORD[48+rdi],xmm0 |
| 2877 | $L$sqr8x_zero_start: |
| 2878 | movdqa XMMWORD[64+rdi],xmm0 |
| 2879 | movdqa XMMWORD[80+rdi],xmm0 |
| 2880 | movdqa XMMWORD[96+rdi],xmm0 |
| 2881 | movdqa XMMWORD[112+rdi],xmm0 |
| 2882 | lea rdi,[128+rdi] |
| 2883 | sub r9,64 |
| 2884 | jnz NEAR $L$sqrx8x_zero |
| 2885 | |
| 2886 | mov rdx,QWORD[rsi] |
| 2887 | |
| 2888 | xor r10,r10 |
| 2889 | xor r11,r11 |
| 2890 | xor r12,r12 |
| 2891 | xor r13,r13 |
| 2892 | xor r14,r14 |
| 2893 | xor r15,r15 |
| 2894 | lea rdi,[((48+8))+rsp] |
| 2895 | xor rbp,rbp |
| 2896 | jmp NEAR $L$sqrx8x_outer_loop |
| 2897 | |
| 2898 | ALIGN 32 |
| 2899 | $L$sqrx8x_outer_loop: |
| 2900 | mulx rax,r8,QWORD[8+rsi] |
| 2901 | adcx r8,r9 |
| 2902 | adox r10,rax |
| 2903 | mulx rax,r9,QWORD[16+rsi] |
| 2904 | adcx r9,r10 |
| 2905 | adox r11,rax |
| 2906 | DB 0xc4,0xe2,0xab,0xf6,0x86,0x18,0x00,0x00,0x00 |
| 2907 | adcx r10,r11 |
| 2908 | adox r12,rax |
| 2909 | DB 0xc4,0xe2,0xa3,0xf6,0x86,0x20,0x00,0x00,0x00 |
| 2910 | adcx r11,r12 |
| 2911 | adox r13,rax |
| 2912 | mulx rax,r12,QWORD[40+rsi] |
| 2913 | adcx r12,r13 |
| 2914 | adox r14,rax |
| 2915 | mulx rax,r13,QWORD[48+rsi] |
| 2916 | adcx r13,r14 |
| 2917 | adox rax,r15 |
| 2918 | mulx r15,r14,QWORD[56+rsi] |
| 2919 | mov rdx,QWORD[8+rsi] |
| 2920 | adcx r14,rax |
| 2921 | adox r15,rbp |
| 2922 | adc r15,QWORD[64+rdi] |
| 2923 | mov QWORD[8+rdi],r8 |
| 2924 | mov QWORD[16+rdi],r9 |
| 2925 | sbb rcx,rcx |
| 2926 | xor rbp,rbp |
| 2927 | |
| 2928 | |
| 2929 | mulx rbx,r8,QWORD[16+rsi] |
| 2930 | mulx rax,r9,QWORD[24+rsi] |
| 2931 | adcx r8,r10 |
| 2932 | adox r9,rbx |
| 2933 | mulx rbx,r10,QWORD[32+rsi] |
| 2934 | adcx r9,r11 |
| 2935 | adox r10,rax |
| 2936 | DB 0xc4,0xe2,0xa3,0xf6,0x86,0x28,0x00,0x00,0x00 |
| 2937 | adcx r10,r12 |
| 2938 | adox r11,rbx |
| 2939 | DB 0xc4,0xe2,0x9b,0xf6,0x9e,0x30,0x00,0x00,0x00 |
| 2940 | adcx r11,r13 |
| 2941 | adox r12,r14 |
| 2942 | DB 0xc4,0x62,0x93,0xf6,0xb6,0x38,0x00,0x00,0x00 |
| 2943 | mov rdx,QWORD[16+rsi] |
| 2944 | adcx r12,rax |
| 2945 | adox r13,rbx |
| 2946 | adcx r13,r15 |
| 2947 | adox r14,rbp |
| 2948 | adcx r14,rbp |
| 2949 | |
| 2950 | mov QWORD[24+rdi],r8 |
| 2951 | mov QWORD[32+rdi],r9 |
| 2952 | |
| 2953 | mulx rbx,r8,QWORD[24+rsi] |
| 2954 | mulx rax,r9,QWORD[32+rsi] |
| 2955 | adcx r8,r10 |
| 2956 | adox r9,rbx |
| 2957 | mulx rbx,r10,QWORD[40+rsi] |
| 2958 | adcx r9,r11 |
| 2959 | adox r10,rax |
| 2960 | DB 0xc4,0xe2,0xa3,0xf6,0x86,0x30,0x00,0x00,0x00 |
| 2961 | adcx r10,r12 |
| 2962 | adox r11,r13 |
| 2963 | DB 0xc4,0x62,0x9b,0xf6,0xae,0x38,0x00,0x00,0x00 |
| 2964 | DB 0x3e |
| 2965 | mov rdx,QWORD[24+rsi] |
| 2966 | adcx r11,rbx |
| 2967 | adox r12,rax |
| 2968 | adcx r12,r14 |
| 2969 | mov QWORD[40+rdi],r8 |
| 2970 | mov QWORD[48+rdi],r9 |
| 2971 | mulx rax,r8,QWORD[32+rsi] |
| 2972 | adox r13,rbp |
| 2973 | adcx r13,rbp |
| 2974 | |
| 2975 | mulx rbx,r9,QWORD[40+rsi] |
| 2976 | adcx r8,r10 |
| 2977 | adox r9,rax |
| 2978 | mulx rax,r10,QWORD[48+rsi] |
| 2979 | adcx r9,r11 |
| 2980 | adox r10,r12 |
| 2981 | mulx r12,r11,QWORD[56+rsi] |
| 2982 | mov rdx,QWORD[32+rsi] |
| 2983 | mov r14,QWORD[40+rsi] |
| 2984 | adcx r10,rbx |
| 2985 | adox r11,rax |
| 2986 | mov r15,QWORD[48+rsi] |
| 2987 | adcx r11,r13 |
| 2988 | adox r12,rbp |
| 2989 | adcx r12,rbp |
| 2990 | |
| 2991 | mov QWORD[56+rdi],r8 |
| 2992 | mov QWORD[64+rdi],r9 |
| 2993 | |
| 2994 | mulx rax,r9,r14 |
| 2995 | mov r8,QWORD[56+rsi] |
| 2996 | adcx r9,r10 |
| 2997 | mulx rbx,r10,r15 |
| 2998 | adox r10,rax |
| 2999 | adcx r10,r11 |
| 3000 | mulx rax,r11,r8 |
| 3001 | mov rdx,r14 |
| 3002 | adox r11,rbx |
| 3003 | adcx r11,r12 |
| 3004 | |
| 3005 | adcx rax,rbp |
| 3006 | |
| 3007 | mulx rbx,r14,r15 |
| 3008 | mulx r13,r12,r8 |
| 3009 | mov rdx,r15 |
| 3010 | lea rsi,[64+rsi] |
| 3011 | adcx r11,r14 |
| 3012 | adox r12,rbx |
| 3013 | adcx r12,rax |
| 3014 | adox r13,rbp |
| 3015 | |
| 3016 | DB 0x67,0x67 |
| 3017 | mulx r14,r8,r8 |
| 3018 | adcx r13,r8 |
| 3019 | adcx r14,rbp |
| 3020 | |
| 3021 | cmp rsi,QWORD[((8+8))+rsp] |
| 3022 | je NEAR $L$sqrx8x_outer_break |
| 3023 | |
| 3024 | neg rcx |
| 3025 | mov rcx,-8 |
| 3026 | mov r15,rbp |
| 3027 | mov r8,QWORD[64+rdi] |
| 3028 | adcx r9,QWORD[72+rdi] |
| 3029 | adcx r10,QWORD[80+rdi] |
| 3030 | adcx r11,QWORD[88+rdi] |
| 3031 | adc r12,QWORD[96+rdi] |
| 3032 | adc r13,QWORD[104+rdi] |
| 3033 | adc r14,QWORD[112+rdi] |
| 3034 | adc r15,QWORD[120+rdi] |
| 3035 | lea rbp,[rsi] |
| 3036 | lea rdi,[128+rdi] |
| 3037 | sbb rax,rax |
| 3038 | |
| 3039 | mov rdx,QWORD[((-64))+rsi] |
| 3040 | mov QWORD[((16+8))+rsp],rax |
| 3041 | mov QWORD[((24+8))+rsp],rdi |
| 3042 | |
| 3043 | |
| 3044 | xor eax,eax |
| 3045 | jmp NEAR $L$sqrx8x_loop |
| 3046 | |
| 3047 | ALIGN 32 |
| 3048 | $L$sqrx8x_loop: |
| 3049 | mov rbx,r8 |
| 3050 | mulx r8,rax,QWORD[rbp] |
| 3051 | adcx rbx,rax |
| 3052 | adox r8,r9 |
| 3053 | |
| 3054 | mulx r9,rax,QWORD[8+rbp] |
| 3055 | adcx r8,rax |
| 3056 | adox r9,r10 |
| 3057 | |
| 3058 | mulx r10,rax,QWORD[16+rbp] |
| 3059 | adcx r9,rax |
| 3060 | adox r10,r11 |
| 3061 | |
| 3062 | mulx r11,rax,QWORD[24+rbp] |
| 3063 | adcx r10,rax |
| 3064 | adox r11,r12 |
| 3065 | |
| 3066 | DB 0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00 |
| 3067 | adcx r11,rax |
| 3068 | adox r12,r13 |
| 3069 | |
| 3070 | mulx r13,rax,QWORD[40+rbp] |
| 3071 | adcx r12,rax |
| 3072 | adox r13,r14 |
| 3073 | |
| 3074 | mulx r14,rax,QWORD[48+rbp] |
| 3075 | mov QWORD[rcx*8+rdi],rbx |
| 3076 | mov ebx,0 |
| 3077 | adcx r13,rax |
| 3078 | adox r14,r15 |
| 3079 | |
| 3080 | DB 0xc4,0x62,0xfb,0xf6,0xbd,0x38,0x00,0x00,0x00 |
| 3081 | mov rdx,QWORD[8+rcx*8+rsi] |
| 3082 | adcx r14,rax |
| 3083 | adox r15,rbx |
| 3084 | adcx r15,rbx |
| 3085 | |
| 3086 | DB 0x67 |
| 3087 | inc rcx |
| 3088 | jnz NEAR $L$sqrx8x_loop |
| 3089 | |
| 3090 | lea rbp,[64+rbp] |
| 3091 | mov rcx,-8 |
| 3092 | cmp rbp,QWORD[((8+8))+rsp] |
| 3093 | je NEAR $L$sqrx8x_break |
| 3094 | |
| 3095 | sub rbx,QWORD[((16+8))+rsp] |
| 3096 | DB 0x66 |
| 3097 | mov rdx,QWORD[((-64))+rsi] |
| 3098 | adcx r8,QWORD[rdi] |
| 3099 | adcx r9,QWORD[8+rdi] |
| 3100 | adc r10,QWORD[16+rdi] |
| 3101 | adc r11,QWORD[24+rdi] |
| 3102 | adc r12,QWORD[32+rdi] |
| 3103 | adc r13,QWORD[40+rdi] |
| 3104 | adc r14,QWORD[48+rdi] |
| 3105 | adc r15,QWORD[56+rdi] |
| 3106 | lea rdi,[64+rdi] |
| 3107 | DB 0x67 |
| 3108 | sbb rax,rax |
| 3109 | xor ebx,ebx |
| 3110 | mov QWORD[((16+8))+rsp],rax |
| 3111 | jmp NEAR $L$sqrx8x_loop |
| 3112 | |
| 3113 | ALIGN 32 |
| 3114 | $L$sqrx8x_break: |
| 3115 | xor rbp,rbp |
| 3116 | sub rbx,QWORD[((16+8))+rsp] |
| 3117 | adcx r8,rbp |
| 3118 | mov rcx,QWORD[((24+8))+rsp] |
| 3119 | adcx r9,rbp |
| 3120 | mov rdx,QWORD[rsi] |
| 3121 | adc r10,0 |
| 3122 | mov QWORD[rdi],r8 |
| 3123 | adc r11,0 |
| 3124 | adc r12,0 |
| 3125 | adc r13,0 |
| 3126 | adc r14,0 |
| 3127 | adc r15,0 |
| 3128 | cmp rdi,rcx |
| 3129 | je NEAR $L$sqrx8x_outer_loop |
| 3130 | |
| 3131 | mov QWORD[8+rdi],r9 |
| 3132 | mov r9,QWORD[8+rcx] |
| 3133 | mov QWORD[16+rdi],r10 |
| 3134 | mov r10,QWORD[16+rcx] |
| 3135 | mov QWORD[24+rdi],r11 |
| 3136 | mov r11,QWORD[24+rcx] |
| 3137 | mov QWORD[32+rdi],r12 |
| 3138 | mov r12,QWORD[32+rcx] |
| 3139 | mov QWORD[40+rdi],r13 |
| 3140 | mov r13,QWORD[40+rcx] |
| 3141 | mov QWORD[48+rdi],r14 |
| 3142 | mov r14,QWORD[48+rcx] |
| 3143 | mov QWORD[56+rdi],r15 |
| 3144 | mov r15,QWORD[56+rcx] |
| 3145 | mov rdi,rcx |
| 3146 | jmp NEAR $L$sqrx8x_outer_loop |
| 3147 | |
| 3148 | ALIGN 32 |
| 3149 | $L$sqrx8x_outer_break: |
| 3150 | mov QWORD[72+rdi],r9 |
| 3151 | DB 102,72,15,126,217 |
| 3152 | mov QWORD[80+rdi],r10 |
| 3153 | mov QWORD[88+rdi],r11 |
| 3154 | mov QWORD[96+rdi],r12 |
| 3155 | mov QWORD[104+rdi],r13 |
| 3156 | mov QWORD[112+rdi],r14 |
| 3157 | lea rdi,[((48+8))+rsp] |
| 3158 | mov rdx,QWORD[rcx*1+rsi] |
| 3159 | |
| 3160 | mov r11,QWORD[8+rdi] |
| 3161 | xor r10,r10 |
| 3162 | mov r9,QWORD[((0+8))+rsp] |
| 3163 | adox r11,r11 |
| 3164 | mov r12,QWORD[16+rdi] |
| 3165 | mov r13,QWORD[24+rdi] |
| 3166 | |
| 3167 | |
| 3168 | ALIGN 32 |
| 3169 | $L$sqrx4x_shift_n_add: |
| 3170 | mulx rbx,rax,rdx |
| 3171 | adox r12,r12 |
| 3172 | adcx rax,r10 |
| 3173 | DB 0x48,0x8b,0x94,0x0e,0x08,0x00,0x00,0x00 |
| 3174 | DB 0x4c,0x8b,0x97,0x20,0x00,0x00,0x00 |
| 3175 | adox r13,r13 |
| 3176 | adcx rbx,r11 |
| 3177 | mov r11,QWORD[40+rdi] |
| 3178 | mov QWORD[rdi],rax |
| 3179 | mov QWORD[8+rdi],rbx |
| 3180 | |
| 3181 | mulx rbx,rax,rdx |
| 3182 | adox r10,r10 |
| 3183 | adcx rax,r12 |
| 3184 | mov rdx,QWORD[16+rcx*1+rsi] |
| 3185 | mov r12,QWORD[48+rdi] |
| 3186 | adox r11,r11 |
| 3187 | adcx rbx,r13 |
| 3188 | mov r13,QWORD[56+rdi] |
| 3189 | mov QWORD[16+rdi],rax |
| 3190 | mov QWORD[24+rdi],rbx |
| 3191 | |
| 3192 | mulx rbx,rax,rdx |
| 3193 | adox r12,r12 |
| 3194 | adcx rax,r10 |
| 3195 | mov rdx,QWORD[24+rcx*1+rsi] |
| 3196 | lea rcx,[32+rcx] |
| 3197 | mov r10,QWORD[64+rdi] |
| 3198 | adox r13,r13 |
| 3199 | adcx rbx,r11 |
| 3200 | mov r11,QWORD[72+rdi] |
| 3201 | mov QWORD[32+rdi],rax |
| 3202 | mov QWORD[40+rdi],rbx |
| 3203 | |
| 3204 | mulx rbx,rax,rdx |
| 3205 | adox r10,r10 |
| 3206 | adcx rax,r12 |
| 3207 | jrcxz $L$sqrx4x_shift_n_add_break |
| 3208 | DB 0x48,0x8b,0x94,0x0e,0x00,0x00,0x00,0x00 |
| 3209 | adox r11,r11 |
| 3210 | adcx rbx,r13 |
| 3211 | mov r12,QWORD[80+rdi] |
| 3212 | mov r13,QWORD[88+rdi] |
| 3213 | mov QWORD[48+rdi],rax |
| 3214 | mov QWORD[56+rdi],rbx |
| 3215 | lea rdi,[64+rdi] |
| 3216 | nop |
| 3217 | jmp NEAR $L$sqrx4x_shift_n_add |
| 3218 | |
| 3219 | ALIGN 32 |
| 3220 | $L$sqrx4x_shift_n_add_break: |
| 3221 | adcx rbx,r13 |
| 3222 | mov QWORD[48+rdi],rax |
| 3223 | mov QWORD[56+rdi],rbx |
| 3224 | lea rdi,[64+rdi] |
| 3225 | DB 102,72,15,126,213 |
| 3226 | __bn_sqrx8x_reduction: |
| 3227 | xor eax,eax |
| 3228 | mov rbx,QWORD[((32+8))+rsp] |
| 3229 | mov rdx,QWORD[((48+8))+rsp] |
| 3230 | lea rcx,[((-64))+r9*1+rbp] |
| 3231 | |
| 3232 | mov QWORD[((0+8))+rsp],rcx |
| 3233 | mov QWORD[((8+8))+rsp],rdi |
| 3234 | |
| 3235 | lea rdi,[((48+8))+rsp] |
| 3236 | jmp NEAR $L$sqrx8x_reduction_loop |
| 3237 | |
| 3238 | ALIGN 32 |
| 3239 | $L$sqrx8x_reduction_loop: |
| 3240 | mov r9,QWORD[8+rdi] |
| 3241 | mov r10,QWORD[16+rdi] |
| 3242 | mov r11,QWORD[24+rdi] |
| 3243 | mov r12,QWORD[32+rdi] |
| 3244 | mov r8,rdx |
| 3245 | imul rdx,rbx |
| 3246 | mov r13,QWORD[40+rdi] |
| 3247 | mov r14,QWORD[48+rdi] |
| 3248 | mov r15,QWORD[56+rdi] |
| 3249 | mov QWORD[((24+8))+rsp],rax |
| 3250 | |
| 3251 | lea rdi,[64+rdi] |
| 3252 | xor rsi,rsi |
| 3253 | mov rcx,-8 |
| 3254 | jmp NEAR $L$sqrx8x_reduce |
| 3255 | |
| 3256 | ALIGN 32 |
| 3257 | $L$sqrx8x_reduce: |
| 3258 | mov rbx,r8 |
| 3259 | mulx r8,rax,QWORD[rbp] |
| 3260 | adcx rax,rbx |
| 3261 | adox r8,r9 |
| 3262 | |
| 3263 | mulx r9,rbx,QWORD[8+rbp] |
| 3264 | adcx r8,rbx |
| 3265 | adox r9,r10 |
| 3266 | |
| 3267 | mulx r10,rbx,QWORD[16+rbp] |
| 3268 | adcx r9,rbx |
| 3269 | adox r10,r11 |
| 3270 | |
| 3271 | mulx r11,rbx,QWORD[24+rbp] |
| 3272 | adcx r10,rbx |
| 3273 | adox r11,r12 |
| 3274 | |
| 3275 | DB 0xc4,0x62,0xe3,0xf6,0xa5,0x20,0x00,0x00,0x00 |
| 3276 | mov rax,rdx |
| 3277 | mov rdx,r8 |
| 3278 | adcx r11,rbx |
| 3279 | adox r12,r13 |
| 3280 | |
| 3281 | mulx rdx,rbx,QWORD[((32+8))+rsp] |
| 3282 | mov rdx,rax |
| 3283 | mov QWORD[((64+48+8))+rcx*8+rsp],rax |
| 3284 | |
| 3285 | mulx r13,rax,QWORD[40+rbp] |
| 3286 | adcx r12,rax |
| 3287 | adox r13,r14 |
| 3288 | |
| 3289 | mulx r14,rax,QWORD[48+rbp] |
| 3290 | adcx r13,rax |
| 3291 | adox r14,r15 |
| 3292 | |
| 3293 | mulx r15,rax,QWORD[56+rbp] |
| 3294 | mov rdx,rbx |
| 3295 | adcx r14,rax |
| 3296 | adox r15,rsi |
| 3297 | adcx r15,rsi |
| 3298 | |
| 3299 | DB 0x67,0x67,0x67 |
| 3300 | inc rcx |
| 3301 | jnz NEAR $L$sqrx8x_reduce |
| 3302 | |
| 3303 | mov rax,rsi |
| 3304 | cmp rbp,QWORD[((0+8))+rsp] |
| 3305 | jae NEAR $L$sqrx8x_no_tail |
| 3306 | |
| 3307 | mov rdx,QWORD[((48+8))+rsp] |
| 3308 | add r8,QWORD[rdi] |
| 3309 | lea rbp,[64+rbp] |
| 3310 | mov rcx,-8 |
| 3311 | adcx r9,QWORD[8+rdi] |
| 3312 | adcx r10,QWORD[16+rdi] |
| 3313 | adc r11,QWORD[24+rdi] |
| 3314 | adc r12,QWORD[32+rdi] |
| 3315 | adc r13,QWORD[40+rdi] |
| 3316 | adc r14,QWORD[48+rdi] |
| 3317 | adc r15,QWORD[56+rdi] |
| 3318 | lea rdi,[64+rdi] |
| 3319 | sbb rax,rax |
| 3320 | |
| 3321 | xor rsi,rsi |
| 3322 | mov QWORD[((16+8))+rsp],rax |
| 3323 | jmp NEAR $L$sqrx8x_tail |
| 3324 | |
| 3325 | ALIGN 32 |
| 3326 | $L$sqrx8x_tail: |
| 3327 | mov rbx,r8 |
| 3328 | mulx r8,rax,QWORD[rbp] |
| 3329 | adcx rbx,rax |
| 3330 | adox r8,r9 |
| 3331 | |
| 3332 | mulx r9,rax,QWORD[8+rbp] |
| 3333 | adcx r8,rax |
| 3334 | adox r9,r10 |
| 3335 | |
| 3336 | mulx r10,rax,QWORD[16+rbp] |
| 3337 | adcx r9,rax |
| 3338 | adox r10,r11 |
| 3339 | |
| 3340 | mulx r11,rax,QWORD[24+rbp] |
| 3341 | adcx r10,rax |
| 3342 | adox r11,r12 |
| 3343 | |
| 3344 | DB 0xc4,0x62,0xfb,0xf6,0xa5,0x20,0x00,0x00,0x00 |
| 3345 | adcx r11,rax |
| 3346 | adox r12,r13 |
| 3347 | |
| 3348 | mulx r13,rax,QWORD[40+rbp] |
| 3349 | adcx r12,rax |
| 3350 | adox r13,r14 |
| 3351 | |
| 3352 | mulx r14,rax,QWORD[48+rbp] |
| 3353 | adcx r13,rax |
| 3354 | adox r14,r15 |
| 3355 | |
| 3356 | mulx r15,rax,QWORD[56+rbp] |
| 3357 | mov rdx,QWORD[((72+48+8))+rcx*8+rsp] |
| 3358 | adcx r14,rax |
| 3359 | adox r15,rsi |
| 3360 | mov QWORD[rcx*8+rdi],rbx |
| 3361 | mov rbx,r8 |
| 3362 | adcx r15,rsi |
| 3363 | |
| 3364 | inc rcx |
| 3365 | jnz NEAR $L$sqrx8x_tail |
| 3366 | |
| 3367 | cmp rbp,QWORD[((0+8))+rsp] |
| 3368 | jae NEAR $L$sqrx8x_tail_done |
| 3369 | |
| 3370 | sub rsi,QWORD[((16+8))+rsp] |
| 3371 | mov rdx,QWORD[((48+8))+rsp] |
| 3372 | lea rbp,[64+rbp] |
| 3373 | adc r8,QWORD[rdi] |
| 3374 | adc r9,QWORD[8+rdi] |
| 3375 | adc r10,QWORD[16+rdi] |
| 3376 | adc r11,QWORD[24+rdi] |
| 3377 | adc r12,QWORD[32+rdi] |
| 3378 | adc r13,QWORD[40+rdi] |
| 3379 | adc r14,QWORD[48+rdi] |
| 3380 | adc r15,QWORD[56+rdi] |
| 3381 | lea rdi,[64+rdi] |
| 3382 | sbb rax,rax |
| 3383 | sub rcx,8 |
| 3384 | |
| 3385 | xor rsi,rsi |
| 3386 | mov QWORD[((16+8))+rsp],rax |
| 3387 | jmp NEAR $L$sqrx8x_tail |
| 3388 | |
| 3389 | ALIGN 32 |
| 3390 | $L$sqrx8x_tail_done: |
| 3391 | xor rax,rax |
| 3392 | add r8,QWORD[((24+8))+rsp] |
| 3393 | adc r9,0 |
| 3394 | adc r10,0 |
| 3395 | adc r11,0 |
| 3396 | adc r12,0 |
| 3397 | adc r13,0 |
| 3398 | adc r14,0 |
| 3399 | adc r15,0 |
| 3400 | adc rax,0 |
| 3401 | |
| 3402 | sub rsi,QWORD[((16+8))+rsp] |
| 3403 | $L$sqrx8x_no_tail: |
| 3404 | adc r8,QWORD[rdi] |
| 3405 | DB 102,72,15,126,217 |
| 3406 | adc r9,QWORD[8+rdi] |
| 3407 | mov rsi,QWORD[56+rbp] |
| 3408 | DB 102,72,15,126,213 |
| 3409 | adc r10,QWORD[16+rdi] |
| 3410 | adc r11,QWORD[24+rdi] |
| 3411 | adc r12,QWORD[32+rdi] |
| 3412 | adc r13,QWORD[40+rdi] |
| 3413 | adc r14,QWORD[48+rdi] |
| 3414 | adc r15,QWORD[56+rdi] |
| 3415 | adc rax,0 |
| 3416 | |
| 3417 | mov rbx,QWORD[((32+8))+rsp] |
| 3418 | mov rdx,QWORD[64+rcx*1+rdi] |
| 3419 | |
| 3420 | mov QWORD[rdi],r8 |
| 3421 | lea r8,[64+rdi] |
| 3422 | mov QWORD[8+rdi],r9 |
| 3423 | mov QWORD[16+rdi],r10 |
| 3424 | mov QWORD[24+rdi],r11 |
| 3425 | mov QWORD[32+rdi],r12 |
| 3426 | mov QWORD[40+rdi],r13 |
| 3427 | mov QWORD[48+rdi],r14 |
| 3428 | mov QWORD[56+rdi],r15 |
| 3429 | |
| 3430 | lea rdi,[64+rcx*1+rdi] |
| 3431 | cmp r8,QWORD[((8+8))+rsp] |
| 3432 | jb NEAR $L$sqrx8x_reduction_loop |
| 3433 | ret |
| 3434 | |
| 3435 | |
| 3436 | ALIGN 32 |
| 3437 | |
| 3438 | __bn_postx4x_internal: |
| 3439 | |
| 3440 | mov r12,QWORD[rbp] |
| 3441 | mov r10,rcx |
| 3442 | mov r9,rcx |
| 3443 | neg rax |
| 3444 | sar rcx,3+2 |
| 3445 | |
| 3446 | DB 102,72,15,126,202 |
| 3447 | DB 102,72,15,126,206 |
| 3448 | dec r12 |
| 3449 | mov r13,QWORD[8+rbp] |
| 3450 | xor r8,r8 |
| 3451 | mov r14,QWORD[16+rbp] |
| 3452 | mov r15,QWORD[24+rbp] |
| 3453 | jmp NEAR $L$sqrx4x_sub_entry |
| 3454 | |
| 3455 | ALIGN 16 |
| 3456 | $L$sqrx4x_sub: |
| 3457 | mov r12,QWORD[rbp] |
| 3458 | mov r13,QWORD[8+rbp] |
| 3459 | mov r14,QWORD[16+rbp] |
| 3460 | mov r15,QWORD[24+rbp] |
| 3461 | $L$sqrx4x_sub_entry: |
| 3462 | andn r12,r12,rax |
| 3463 | lea rbp,[32+rbp] |
| 3464 | andn r13,r13,rax |
| 3465 | andn r14,r14,rax |
| 3466 | andn r15,r15,rax |
| 3467 | |
| 3468 | neg r8 |
| 3469 | adc r12,QWORD[rdi] |
| 3470 | adc r13,QWORD[8+rdi] |
| 3471 | adc r14,QWORD[16+rdi] |
| 3472 | adc r15,QWORD[24+rdi] |
| 3473 | mov QWORD[rdx],r12 |
| 3474 | lea rdi,[32+rdi] |
| 3475 | mov QWORD[8+rdx],r13 |
| 3476 | sbb r8,r8 |
| 3477 | mov QWORD[16+rdx],r14 |
| 3478 | mov QWORD[24+rdx],r15 |
| 3479 | lea rdx,[32+rdx] |
| 3480 | |
| 3481 | inc rcx |
| 3482 | jnz NEAR $L$sqrx4x_sub |
| 3483 | |
| 3484 | neg r9 |
| 3485 | |
| 3486 | ret |
| 3487 | |
| 3488 | |
| 3489 | global bn_scatter5 |
| 3490 | |
| 3491 | ALIGN 16 |
| 3492 | bn_scatter5: |
| 3493 | |
| 3494 | _CET_ENDBR |
| 3495 | cmp edx,0 |
| 3496 | jz NEAR $L$scatter_epilogue |
| 3497 | |
| 3498 | |
| 3499 | |
| 3500 | |
| 3501 | |
| 3502 | |
| 3503 | |
| 3504 | |
| 3505 | |
| 3506 | lea r8,[r9*8+r8] |
| 3507 | $L$scatter: |
| 3508 | mov rax,QWORD[rcx] |
| 3509 | lea rcx,[8+rcx] |
| 3510 | mov QWORD[r8],rax |
| 3511 | lea r8,[256+r8] |
| 3512 | sub edx,1 |
| 3513 | jnz NEAR $L$scatter |
| 3514 | $L$scatter_epilogue: |
| 3515 | ret |
| 3516 | |
| 3517 | |
| 3518 | |
| 3519 | global bn_gather5 |
| 3520 | |
| 3521 | ALIGN 32 |
| 3522 | bn_gather5: |
| 3523 | |
| 3524 | $L$SEH_begin_bn_gather5: |
| 3525 | _CET_ENDBR |
| 3526 | |
| 3527 | DB 0x4c,0x8d,0x14,0x24 |
| 3528 | |
| 3529 | DB 0x48,0x81,0xec,0x08,0x01,0x00,0x00 |
| 3530 | lea rax,[$L$inc] |
| 3531 | and rsp,-16 |
| 3532 | |
| 3533 | movd xmm5,r9d |
| 3534 | movdqa xmm0,XMMWORD[rax] |
| 3535 | movdqa xmm1,XMMWORD[16+rax] |
| 3536 | lea r11,[128+r8] |
| 3537 | lea rax,[128+rsp] |
| 3538 | |
| 3539 | pshufd xmm5,xmm5,0 |
| 3540 | movdqa xmm4,xmm1 |
| 3541 | movdqa xmm2,xmm1 |
| 3542 | paddd xmm1,xmm0 |
| 3543 | pcmpeqd xmm0,xmm5 |
| 3544 | movdqa xmm3,xmm4 |
| 3545 | |
| 3546 | paddd xmm2,xmm1 |
| 3547 | pcmpeqd xmm1,xmm5 |
| 3548 | movdqa XMMWORD[(-128)+rax],xmm0 |
| 3549 | movdqa xmm0,xmm4 |
| 3550 | |
| 3551 | paddd xmm3,xmm2 |
| 3552 | pcmpeqd xmm2,xmm5 |
| 3553 | movdqa XMMWORD[(-112)+rax],xmm1 |
| 3554 | movdqa xmm1,xmm4 |
| 3555 | |
| 3556 | paddd xmm0,xmm3 |
| 3557 | pcmpeqd xmm3,xmm5 |
| 3558 | movdqa XMMWORD[(-96)+rax],xmm2 |
| 3559 | movdqa xmm2,xmm4 |
| 3560 | paddd xmm1,xmm0 |
| 3561 | pcmpeqd xmm0,xmm5 |
| 3562 | movdqa XMMWORD[(-80)+rax],xmm3 |
| 3563 | movdqa xmm3,xmm4 |
| 3564 | |
| 3565 | paddd xmm2,xmm1 |
| 3566 | pcmpeqd xmm1,xmm5 |
| 3567 | movdqa XMMWORD[(-64)+rax],xmm0 |
| 3568 | movdqa xmm0,xmm4 |
| 3569 | |
| 3570 | paddd xmm3,xmm2 |
| 3571 | pcmpeqd xmm2,xmm5 |
| 3572 | movdqa XMMWORD[(-48)+rax],xmm1 |
| 3573 | movdqa xmm1,xmm4 |
| 3574 | |
| 3575 | paddd xmm0,xmm3 |
| 3576 | pcmpeqd xmm3,xmm5 |
| 3577 | movdqa XMMWORD[(-32)+rax],xmm2 |
| 3578 | movdqa xmm2,xmm4 |
| 3579 | paddd xmm1,xmm0 |
| 3580 | pcmpeqd xmm0,xmm5 |
| 3581 | movdqa XMMWORD[(-16)+rax],xmm3 |
| 3582 | movdqa xmm3,xmm4 |
| 3583 | |
| 3584 | paddd xmm2,xmm1 |
| 3585 | pcmpeqd xmm1,xmm5 |
| 3586 | movdqa XMMWORD[rax],xmm0 |
| 3587 | movdqa xmm0,xmm4 |
| 3588 | |
| 3589 | paddd xmm3,xmm2 |
| 3590 | pcmpeqd xmm2,xmm5 |
| 3591 | movdqa XMMWORD[16+rax],xmm1 |
| 3592 | movdqa xmm1,xmm4 |
| 3593 | |
| 3594 | paddd xmm0,xmm3 |
| 3595 | pcmpeqd xmm3,xmm5 |
| 3596 | movdqa XMMWORD[32+rax],xmm2 |
| 3597 | movdqa xmm2,xmm4 |
| 3598 | paddd xmm1,xmm0 |
| 3599 | pcmpeqd xmm0,xmm5 |
| 3600 | movdqa XMMWORD[48+rax],xmm3 |
| 3601 | movdqa xmm3,xmm4 |
| 3602 | |
| 3603 | paddd xmm2,xmm1 |
| 3604 | pcmpeqd xmm1,xmm5 |
| 3605 | movdqa XMMWORD[64+rax],xmm0 |
| 3606 | movdqa xmm0,xmm4 |
| 3607 | |
| 3608 | paddd xmm3,xmm2 |
| 3609 | pcmpeqd xmm2,xmm5 |
| 3610 | movdqa XMMWORD[80+rax],xmm1 |
| 3611 | movdqa xmm1,xmm4 |
| 3612 | |
| 3613 | paddd xmm0,xmm3 |
| 3614 | pcmpeqd xmm3,xmm5 |
| 3615 | movdqa XMMWORD[96+rax],xmm2 |
| 3616 | movdqa xmm2,xmm4 |
| 3617 | movdqa XMMWORD[112+rax],xmm3 |
| 3618 | jmp NEAR $L$gather |
| 3619 | |
| 3620 | ALIGN 32 |
| 3621 | $L$gather: |
| 3622 | pxor xmm4,xmm4 |
| 3623 | pxor xmm5,xmm5 |
| 3624 | movdqa xmm0,XMMWORD[((-128))+r11] |
| 3625 | movdqa xmm1,XMMWORD[((-112))+r11] |
| 3626 | movdqa xmm2,XMMWORD[((-96))+r11] |
| 3627 | pand xmm0,XMMWORD[((-128))+rax] |
| 3628 | movdqa xmm3,XMMWORD[((-80))+r11] |
| 3629 | pand xmm1,XMMWORD[((-112))+rax] |
| 3630 | por xmm4,xmm0 |
| 3631 | pand xmm2,XMMWORD[((-96))+rax] |
| 3632 | por xmm5,xmm1 |
| 3633 | pand xmm3,XMMWORD[((-80))+rax] |
| 3634 | por xmm4,xmm2 |
| 3635 | por xmm5,xmm3 |
| 3636 | movdqa xmm0,XMMWORD[((-64))+r11] |
| 3637 | movdqa xmm1,XMMWORD[((-48))+r11] |
| 3638 | movdqa xmm2,XMMWORD[((-32))+r11] |
| 3639 | pand xmm0,XMMWORD[((-64))+rax] |
| 3640 | movdqa xmm3,XMMWORD[((-16))+r11] |
| 3641 | pand xmm1,XMMWORD[((-48))+rax] |
| 3642 | por xmm4,xmm0 |
| 3643 | pand xmm2,XMMWORD[((-32))+rax] |
| 3644 | por xmm5,xmm1 |
| 3645 | pand xmm3,XMMWORD[((-16))+rax] |
| 3646 | por xmm4,xmm2 |
| 3647 | por xmm5,xmm3 |
| 3648 | movdqa xmm0,XMMWORD[r11] |
| 3649 | movdqa xmm1,XMMWORD[16+r11] |
| 3650 | movdqa xmm2,XMMWORD[32+r11] |
| 3651 | pand xmm0,XMMWORD[rax] |
| 3652 | movdqa xmm3,XMMWORD[48+r11] |
| 3653 | pand xmm1,XMMWORD[16+rax] |
| 3654 | por xmm4,xmm0 |
| 3655 | pand xmm2,XMMWORD[32+rax] |
| 3656 | por xmm5,xmm1 |
| 3657 | pand xmm3,XMMWORD[48+rax] |
| 3658 | por xmm4,xmm2 |
| 3659 | por xmm5,xmm3 |
| 3660 | movdqa xmm0,XMMWORD[64+r11] |
| 3661 | movdqa xmm1,XMMWORD[80+r11] |
| 3662 | movdqa xmm2,XMMWORD[96+r11] |
| 3663 | pand xmm0,XMMWORD[64+rax] |
| 3664 | movdqa xmm3,XMMWORD[112+r11] |
| 3665 | pand xmm1,XMMWORD[80+rax] |
| 3666 | por xmm4,xmm0 |
| 3667 | pand xmm2,XMMWORD[96+rax] |
| 3668 | por xmm5,xmm1 |
| 3669 | pand xmm3,XMMWORD[112+rax] |
| 3670 | por xmm4,xmm2 |
| 3671 | por xmm5,xmm3 |
| 3672 | por xmm4,xmm5 |
| 3673 | lea r11,[256+r11] |
| 3674 | |
| 3675 | pshufd xmm0,xmm4,0x4e |
| 3676 | por xmm0,xmm4 |
| 3677 | movq QWORD[rcx],xmm0 |
| 3678 | lea rcx,[8+rcx] |
| 3679 | sub edx,1 |
| 3680 | jnz NEAR $L$gather |
| 3681 | |
| 3682 | lea rsp,[r10] |
| 3683 | |
| 3684 | ret |
| 3685 | $L$SEH_end_bn_gather5: |
| 3686 | |
| 3687 | |
| 3688 | section .rdata rdata align=8 |
| 3689 | ALIGN 64 |
| 3690 | $L$inc: |
| 3691 | DD 0,0,1,1 |
| 3692 | DD 2,2,2,2 |
| 3693 | DB 77,111,110,116,103,111,109,101,114,121,32,77,117,108,116,105 |
| 3694 | DB 112,108,105,99,97,116,105,111,110,32,119,105,116,104,32,115 |
| 3695 | DB 99,97,116,116,101,114,47,103,97,116,104,101,114,32,102,111 |
| 3696 | DB 114,32,120,56,54,95,54,52,44,32,67,82,89,80,84,79 |
| 3697 | DB 71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111 |
| 3698 | DB 112,101,110,115,115,108,46,111,114,103,62,0 |
| 3699 | section .text |
| 3700 | |
| 3701 | EXTERN __imp_RtlVirtualUnwind |
| 3702 | |
| 3703 | ALIGN 16 |
| 3704 | mul_handler: |
| 3705 | push rsi |
| 3706 | push rdi |
| 3707 | push rbx |
| 3708 | push rbp |
| 3709 | push r12 |
| 3710 | push r13 |
| 3711 | push r14 |
| 3712 | push r15 |
| 3713 | pushfq |
| 3714 | sub rsp,64 |
| 3715 | |
| 3716 | mov rax,QWORD[120+r8] |
| 3717 | mov rbx,QWORD[248+r8] |
| 3718 | |
| 3719 | mov rsi,QWORD[8+r9] |
| 3720 | mov r11,QWORD[56+r9] |
| 3721 | |
| 3722 | mov r10d,DWORD[r11] |
| 3723 | lea r10,[r10*1+rsi] |
| 3724 | cmp rbx,r10 |
| 3725 | jb NEAR $L$common_seh_tail |
| 3726 | |
| 3727 | mov r10d,DWORD[4+r11] |
| 3728 | lea r10,[r10*1+rsi] |
| 3729 | cmp rbx,r10 |
| 3730 | jb NEAR $L$common_pop_regs |
| 3731 | |
| 3732 | mov rax,QWORD[152+r8] |
| 3733 | |
| 3734 | mov r10d,DWORD[8+r11] |
| 3735 | lea r10,[r10*1+rsi] |
| 3736 | cmp rbx,r10 |
| 3737 | jae NEAR $L$common_seh_tail |
| 3738 | |
| 3739 | lea r10,[$L$mul_epilogue] |
| 3740 | cmp rbx,r10 |
| 3741 | ja NEAR $L$body_40 |
| 3742 | |
| 3743 | mov r10,QWORD[192+r8] |
| 3744 | mov rax,QWORD[8+r10*8+rax] |
| 3745 | |
| 3746 | jmp NEAR $L$common_pop_regs |
| 3747 | |
| 3748 | $L$body_40: |
| 3749 | mov rax,QWORD[40+rax] |
| 3750 | $L$common_pop_regs: |
| 3751 | mov rbx,QWORD[((-8))+rax] |
| 3752 | mov rbp,QWORD[((-16))+rax] |
| 3753 | mov r12,QWORD[((-24))+rax] |
| 3754 | mov r13,QWORD[((-32))+rax] |
| 3755 | mov r14,QWORD[((-40))+rax] |
| 3756 | mov r15,QWORD[((-48))+rax] |
| 3757 | mov QWORD[144+r8],rbx |
| 3758 | mov QWORD[160+r8],rbp |
| 3759 | mov QWORD[216+r8],r12 |
| 3760 | mov QWORD[224+r8],r13 |
| 3761 | mov QWORD[232+r8],r14 |
| 3762 | mov QWORD[240+r8],r15 |
| 3763 | |
| 3764 | $L$common_seh_tail: |
| 3765 | mov rdi,QWORD[8+rax] |
| 3766 | mov rsi,QWORD[16+rax] |
| 3767 | mov QWORD[152+r8],rax |
| 3768 | mov QWORD[168+r8],rsi |
| 3769 | mov QWORD[176+r8],rdi |
| 3770 | |
| 3771 | mov rdi,QWORD[40+r9] |
| 3772 | mov rsi,r8 |
| 3773 | mov ecx,154 |
| 3774 | DD 0xa548f3fc |
| 3775 | |
| 3776 | mov rsi,r9 |
| 3777 | xor rcx,rcx |
| 3778 | mov rdx,QWORD[8+rsi] |
| 3779 | mov r8,QWORD[rsi] |
| 3780 | mov r9,QWORD[16+rsi] |
| 3781 | mov r10,QWORD[40+rsi] |
| 3782 | lea r11,[56+rsi] |
| 3783 | lea r12,[24+rsi] |
| 3784 | mov QWORD[32+rsp],r10 |
| 3785 | mov QWORD[40+rsp],r11 |
| 3786 | mov QWORD[48+rsp],r12 |
| 3787 | mov QWORD[56+rsp],rcx |
| 3788 | call QWORD[__imp_RtlVirtualUnwind] |
| 3789 | |
| 3790 | mov eax,1 |
| 3791 | add rsp,64 |
| 3792 | popfq |
| 3793 | pop r15 |
| 3794 | pop r14 |
| 3795 | pop r13 |
| 3796 | pop r12 |
| 3797 | pop rbp |
| 3798 | pop rbx |
| 3799 | pop rdi |
| 3800 | pop rsi |
| 3801 | ret |
| 3802 | |
| 3803 | |
| 3804 | section .pdata rdata align=4 |
| 3805 | ALIGN 4 |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 3806 | DD $L$SEH_begin_bn_mul_mont_gather5_nohw wrt ..imagebase |
| 3807 | DD $L$SEH_end_bn_mul_mont_gather5_nohw wrt ..imagebase |
| 3808 | DD $L$SEH_info_bn_mul_mont_gather5_nohw wrt ..imagebase |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 3809 | |
| 3810 | DD $L$SEH_begin_bn_mul4x_mont_gather5 wrt ..imagebase |
| 3811 | DD $L$SEH_end_bn_mul4x_mont_gather5 wrt ..imagebase |
| 3812 | DD $L$SEH_info_bn_mul4x_mont_gather5 wrt ..imagebase |
| 3813 | |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 3814 | DD $L$SEH_begin_bn_power5_nohw wrt ..imagebase |
| 3815 | DD $L$SEH_end_bn_power5_nohw wrt ..imagebase |
| 3816 | DD $L$SEH_info_bn_power5_nohw wrt ..imagebase |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 3817 | DD $L$SEH_begin_bn_mulx4x_mont_gather5 wrt ..imagebase |
| 3818 | DD $L$SEH_end_bn_mulx4x_mont_gather5 wrt ..imagebase |
| 3819 | DD $L$SEH_info_bn_mulx4x_mont_gather5 wrt ..imagebase |
| 3820 | |
| 3821 | DD $L$SEH_begin_bn_powerx5 wrt ..imagebase |
| 3822 | DD $L$SEH_end_bn_powerx5 wrt ..imagebase |
| 3823 | DD $L$SEH_info_bn_powerx5 wrt ..imagebase |
| 3824 | DD $L$SEH_begin_bn_gather5 wrt ..imagebase |
| 3825 | DD $L$SEH_end_bn_gather5 wrt ..imagebase |
| 3826 | DD $L$SEH_info_bn_gather5 wrt ..imagebase |
| 3827 | |
| 3828 | section .xdata rdata align=8 |
| 3829 | ALIGN 8 |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 3830 | $L$SEH_info_bn_mul_mont_gather5_nohw: |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 3831 | DB 9,0,0,0 |
| 3832 | DD mul_handler wrt ..imagebase |
| 3833 | DD $L$mul_body wrt ..imagebase,$L$mul_body wrt ..imagebase,$L$mul_epilogue wrt ..imagebase |
| 3834 | ALIGN 8 |
| 3835 | $L$SEH_info_bn_mul4x_mont_gather5: |
| 3836 | DB 9,0,0,0 |
| 3837 | DD mul_handler wrt ..imagebase |
| 3838 | DD $L$mul4x_prologue wrt ..imagebase,$L$mul4x_body wrt ..imagebase,$L$mul4x_epilogue wrt ..imagebase |
| 3839 | ALIGN 8 |
David Benjamin | 3efe2eb | 2024-05-08 22:24:27 -0700 | [diff] [blame] | 3840 | $L$SEH_info_bn_power5_nohw: |
David Benjamin | fe0c91e | 2024-03-18 15:37:24 +1000 | [diff] [blame] | 3841 | DB 9,0,0,0 |
| 3842 | DD mul_handler wrt ..imagebase |
| 3843 | DD $L$power5_prologue wrt ..imagebase,$L$power5_body wrt ..imagebase,$L$power5_epilogue wrt ..imagebase |
| 3844 | ALIGN 8 |
| 3845 | $L$SEH_info_bn_mulx4x_mont_gather5: |
| 3846 | DB 9,0,0,0 |
| 3847 | DD mul_handler wrt ..imagebase |
| 3848 | DD $L$mulx4x_prologue wrt ..imagebase,$L$mulx4x_body wrt ..imagebase,$L$mulx4x_epilogue wrt ..imagebase |
| 3849 | ALIGN 8 |
| 3850 | $L$SEH_info_bn_powerx5: |
| 3851 | DB 9,0,0,0 |
| 3852 | DD mul_handler wrt ..imagebase |
| 3853 | DD $L$powerx5_prologue wrt ..imagebase,$L$powerx5_body wrt ..imagebase,$L$powerx5_epilogue wrt ..imagebase |
| 3854 | ALIGN 8 |
| 3855 | $L$SEH_info_bn_gather5: |
| 3856 | DB 0x01,0x0b,0x03,0x0a |
| 3857 | DB 0x0b,0x01,0x21,0x00 |
| 3858 | DB 0x04,0xa3,0x00,0x00 |
| 3859 | ALIGN 8 |
| 3860 | %else |
| 3861 | ; Work around https://bugzilla.nasm.us/show_bug.cgi?id=3392738 |
| 3862 | ret |
| 3863 | %endif |